Memory / context lens — RUN AGS-20260821-sequentialthinking

Logical record: step 6, embedded memory/context lens. IDs are those registered in evidence-packet.md §4; new records are proposed under the lens-local MEM-* tag in §6 and are not canonical until the orchestrator registers them.

Search boundary for every negative result in this lens is B0 as defined in evidence-packet.md §4e (the four files of src/sequentialthinking at 2ecb382, each read in full, plus targeted symbol reads), unless a narrower boundary is named inline.

ABS-8 applies globally: there is no test, fixture, trace, log, or run artifact anywhere in the boundary. No finding in this lens can reach observed or causally supported. Every status below is implemented, claimed, absent, or uninspected, and the ceiling is restated at each point where it bites.

1. Retained-part inventory

Parts are split wherever content, form, producer/consumer, checks, or authority path differ. The nine required fields are carried across two tables (1a: identity, substrate, form, persistence, producer, consumer; 1b: lineage, invalidation, regeneration, promotion) because a ten-column table is unreadable; the rows are keyed identically.

Classification distinguishes three things the instruction's memory definition keeps apart:

  • accumulated — material accumulated or changed through use, retained inside the boundary;
  • static-shipped — retained material compiled into the artifact, not changed by use (documentation, tool specification): retained state, not memory;
  • emitted — material that leaves the process and is not retained by the server at all.

1a. Identity, substrate, form, persistence, producer, consumer

Part Class Storage substrate Representational form Persistence Producer Consumer
OBJ-3 — validated ThoughtData record, one per accepted call (index.ts:45-55) accumulated process heap (a fresh object literal, not the caller's argument object) structured record, 9 fields process lifetime; no serialization (ABS-2) CMP-6 validateThoughtData, from model-authored request.params.arguments RTE-6 (append), RTE-7 (conditional index), CMP-7 formatThought (index.ts:103, current record only). Per-field consumer split below
OBJ-4-content — the ThoughtData elements held in thoughtHistory (index.ts:26, :94) accumulated process heap, append-only array structured record array process lifetime; unbounded, never pruned RTE-6, unconditional on every valid call none inside the boundary (ABS-3). No operation reads any element
MEM-1 — history cardinality, the accumulated scalar view of OBJ-4 (index.ts:114) accumulated derived on read from OBJ-4.length; not separately stored integer recomputed per read; lives as long as OBJ-4 implicitly by every RTE-6 append RTE-9 → OBJ-6.thoughtHistoryLength → host LLM (BAP-2)
OBJ-5-content — the ThoughtData elements held in branch buckets (index.ts:27, :96-101) accumulated process heap, keyed buckets; elements are the same object references as the corresponding OBJ-4 entries (see MEM-4) keyed structured records process lifetime; never pruned RTE-7, only when branchFromThought && branchId are both truthy none inside the boundary (ABS-3)
MEM-2 — branch-key set, the accumulated key view of OBJ-5 (created index.ts:97-99, read :113) accumulated property keys of the branches object set of caller-authored strings process lifetime; monotonic — a key is never removed RTE-7 lazy bucket creation, keyed by the model-supplied branchId RTE-9 → OBJ-6.branches → host LLM (BAP-2)
OBJ-1 — tool description prose, 54 lines (SRC-1a:135-188) static-shipped string literal compiled into dist/index.js (CMP-9) natural-language shipped-artifact lifetime; invariant under use authors, at build time host LLM via RTE-2/BAP-1; placement into model context is host-owned and uninspected
OBJ-2 — tool input JSON Schema, 9 properties (SRC-1:189-234) static-shipped object literal compiled into dist/index.js symbolic shipped-artifact lifetime; invariant under use authors, at build time host/model at call time via RTE-2; independently re-checked in-process by RTE-4 (BAP-4)
OBJ-9 — README doctrine (SRC-2) static-shipped repo file; not in files: ["dist"] (SRC-3:13-15), so not in the npm artifact natural-language repo lifetime authors human operator (BAP-5)
OBJ-10 — distribution manifest (SRC-3) static-shipped repo file symbolic config repo lifetime authors npm / host launcher (RTE-13)
OBJ-8 — rendered thought box, full thought text (index.ts:58-84, :104) emitted process stderr; the receiving sink is host-owned and excluded from the boundary natural-language display not retained by the server; downstream retention uninspected RTE-8, per call, from the current record only — it never reads the store human operator / log reader (BAP-3); no consumer inside the boundary
OBJ-6 — success payload, 5 fields (index.ts:106-117) emitted MCP response over stdio structured record → JSON text in one text content block one response; server keeps no copy RTE-9 host LLM (BAP-2)
OBJ-7 — error payload (index.ts:118-129) emitted MCP response over stdio structured record → JSON text one response RTE-10 / RTE-11 host LLM. Carries no accumulated state at all — not even thoughtHistoryLength

Per-field consumer split inside OBJ-3. The bundled record's fields differ by consumer, which matters because it determines what the store could ever return:

Field Consumer(s) Anchor
thought formatThought → stderr only (BAP-3) index.ts:59, :76, :82
thoughtNumber, totalThoughts, nextThoughtNeeded stderr render, and echoed into OBJ-6 from the current record, not from the store index.ts:110-112
isRevision, revisesThought stderr render only; no state operation (ABS-6) index.ts:64-66
branchFromThought, branchId index routing (RTE-7) + stderr render index.ts:67-69, :96-100
needsMoreThoughts none anywhere (ABS-4) — retained but unconsumed index.ts:54

This is the precise sense in which OBJ-6's first three fields are echo, not read-back: they are read off validatedInput, the current call's record (index.ts:110-112), never off OBJ-4 or OBJ-5. Only branches and thoughtHistoryLength are functions of accumulated state.

1b. Lineage, invalidation, regeneration, promotion

Part Lineage Invalidation conditions Regeneration conditions Promotion path toward stronger form or force
OBJ-3 model-authored call arguments → lossy 9-field whitelist projection (MEM-3, index.ts:45-55: any key outside the whitelist is dropped) → in-place field overwrite by the RTE-5 clamp at index.ts:90-92, applied before the push. The retained record therefore diverges from what the model submitted in the totalThoughts field, silently and without record of the original value none post-store none absent (B0). No code path converts a record into any other form. JSON.stringify (:109) and formatThought (:78-83) are lossy renderings toward output, not promotions
OBJ-4-content one append per accepted call, in acceptance order, no separators, no session key (singleton at index.ts:249). Rejected calls are not appended (RTE-10), so the array records accepted calls, not attempts absent post-store (B0): no entry is removed, overwritten, marked, superseded, or expired. Whole store is discarded on process exit, including the process.exit(1) path at index.ts:277 absent: a restarted process starts empty; nothing rebuilds it (ABS-2) absent (B0)
MEM-1 derived from OBJ-4 length; monotonically increasing, one per accepted call recomputed on each read recomputed on each read absent
OBJ-5-content the record is stored twice by reference, once in OBJ-4 and once in the bucket (MEM-4). This is an index over OBJ-4, not a second copy absent post-store absent absent
MEM-2 keys are model-authored branchId strings, cast without type check (index.ts:53) and used directly as object keys, so a non-string value is coerced by JS property access. Lazy creation only; no key normalization, validation, or namespacing absent: monotonic, a branch cannot be closed or removed absent absent
OBJ-1, OBJ-2 authored at build time; compiled into dist/index.js by tsc (CMP-9) invalidated only by a new build and publish (RTE-13) regenerated by rebuild not applicable — already the shipped form. OBJ-2 already carries the strongest force in the boundary (BAP-4, enforcing)
OBJ-9, OBJ-10 authored; OBJ-9 is not shipped in the npm artifact new commit rewrite not applicable
OBJ-8 current record → chalk/unicode render; the render reads the record, never the store (index.ts:103) not retained by the server re-emitted only if the same call recurs absent inside the boundary. Whether the host's stderr sink turns it into durable material is uninspected
OBJ-6, OBJ-7 constructed per call and returned — — absent

Nothing in the boundary carries a consolidate, import, merge, or distil label, so there is no lineage or curation label available to be mistaken for semantic preservation. The append at index.ts:94 asserts no transformation and warrants none: it is raw retention of a normalized record.

2. Write side vs read-back

2.1 Write agency

Automatic within the server; agent-triggered at its boundary. Every valid tools/call writes (index.ts:94, unconditional). There is no manual write path, no operator write API, no configuration that could disable, filter, or throttle writing (ABS-1, ABS-2). The content is authored entirely by an external agent, the host LLM, which is excluded from the boundary; the decision to write is not the server's at any point.

There is exactly one gate between an agent's utterance and the store: RTE-4 / BAP-4, four presence-and-type predicates over the envelope (index.ts:32-43). It checks nothing about the substance of a thought (ABS-5). Status: implemented.

2.2 Acquisition, index maintenance, curation — separated

Function Present? Mechanism / evidence
Acquisition implemented RTE-4 envelope gate → RTE-6 unconditional append (index.ts:88-94). No relevance, novelty, or quality condition (ABS-9)
Normalization implemented MEM-3: the whitelist projection at index.ts:45-55 plus the pre-push clamp RTE-5. Lossy in two directions — unknown keys dropped, totalThoughts overwritten
Index maintenance implemented, thin RTE-7 only. A single flat keyed index over caller-supplied branchId, lazy bucket creation, no key hygiene. The gate is branchFromThought && branchId truthiness (index.ts:96), so branchId: "" or branchFromThought: 0 silently skips indexing while the thought still lands in OBJ-4 — a silent index miss with no error and no signal to the caller
Curation absent (B0) See breakdown below

Curation breakdown. Each is absent within B0, and each names what it prevents:

Operation Status Evidence Conclusion prevented
Consolidation absent No operation reads any stored element (ABS-3; only reads are index.ts:113-114) Prevents any claim that stored thoughts are merged, summarized, or compacted
Deduplication absent index.ts:94 pushes with no comparison to any prior entry Prevents any claim that repeated or near-identical thoughts are collapsed
Evolution / revision effect absent ABS-6: isRevision and revisesThought are stored and rendered, but no state operation links, supersedes, marks, or removes a target entry; the revising thought is appended like any other Prevents any claim of a revision graph, supersession, or corrected history. Revision is labelled, not effected
Synthesis absent No read of content exists to synthesize from (ABS-3) Prevents any claim that the server produces a derived or distilled artifact from the store
Invalidation absent No entry is removed, overwritten, or marked after the push. The only field write, RTE-5, happens before storage Prevents any claim of post-hoc correction or retraction
Decay / eviction / bounding absent No TTL, cap, compaction, or eviction; array is append-only and unbounded (runtime §5.4, Performance) Prevents any claim that the store is bounded or self-limiting in a long session
Promotion absent No serialization, file, or export path (ABS-2) Prevents any claim of a path from in-memory record toward a stronger form or force

2.3 Raw traces vs distilled artifacts

There is no distilled tier. The store holds normalized raw records (MEM-3); the only non-raw accumulated artifacts are MEM-1 and MEM-2, which are a cardinality and a key set — aggregate projections, not content distillations. Whether the raw records are traces or working material is not settled inside the boundary, because nothing consumes them.

2.4 Read-back — does the derived return count?

Definition applied as given: memory read-back is material accumulated or changed through use returning to a later invocation or action.

Yes, the derived return at index.ts:113-114 is read-back, and it is the only read-back in the system. Both values are functions of state accumulated across prior invocations, and both are returned into a later invocation's response. Three precisions bound what that means:

  1. It is fused with a same-turn increment. The write (:94) precedes the read (:113-114) inside a single processThought call. The returned thoughtHistoryLength therefore equals (accepted calls before this one) + 1, and a newly created branch key appears in the very response that created it. The accumulated component is real, but it never arrives alone.
  2. What it can carry. thoughtHistoryLength: a count of accepted calls on this process since start — scoped to the process, not the conversation, because the singleton at index.ts:249 has no session key or request partition. branches: the set of distinct caller-authored branch labels ever opened. Note carefully — branches is the one place accumulated, model-authored text returns to the model. This does not contradict ABS-3, which is scoped to stored thought content; branch labels are identifiers stored as keys, and they do come back.
  3. What it cannot carry. No thought text, no thought metadata, no per-entry addressing, no ordering, no branch membership (the response gives the key set only, never which thoughts are in a branch), and no indication that any prior call was rejected other than through the counter.

What a receiving agent could actually do with it (implemented-level reasoning about affordance; ABS-8 prevents any claim that an agent does any of this):

  • confirm the call was accepted at all — an error payload (OBJ-7) omits the counters entirely;
  • recover branch labels it opened earlier and has since lost from its own context — the only genuine recall the channel offers;
  • detect a divergence between its own expected count and thoughtHistoryLength, which would indicate either rejected calls (RTE-10 does not append) or another conversation multiplexed onto the same process.

It cannot retrieve, review, resume, re-read, verify, or reason over any prior thought.

The sharp asymmetry: in the ordinary single-conversation case the model authored every thought in the store, so the projection is largely redundant with the model's own context — the server returns a count of what the caller itself sent. The store's function inside the boundary is therefore closer to a write-only ledger with a receipt than to a memory the agent reads.

2.5 The stderr path, stated exactly (RTE-8, OBJ-8, BAP-3)

Full thought content does leave the process here, on every valid call, unconditionally and untunably (ABS-1). It is the richest output in the system and the only place the thought string exits. Under this lens's definitions:

  • It is not read-back. It is push-emitted at write time toward a human log reader (BAP-3), not returned to a later invocation or action of any agent. Nothing in the boundary consumes it.
  • It is not a store read. formatThought is called on validatedInput (index.ts:103) and reads only the current record (index.ts:59). Even the rich channel performs no read of OBJ-4 or OBJ-5.
  • It is not retained by the server. Whether it accumulates depends on the host's stderr handling, which the boundary excludes and this lens did not inspect.

The system's richest content channel points away from the agent that produced the material and carries display-only force.

3. Context-route annotation

RTE-9 is the read-back route. RTE-2 is annotated as well, because it is the system's other outbound context route and its annotation is load-bearing as a negative: it is where the memory mechanism is not.

RTE-9 — success response construction (index.ts:106-117)

Field Annotation
Read-back direction (receiving agent's perspective) Pull, at the transport level: the host LLM's own tools/call elicits the response, and the server never initiates (no sampling capability, ABS-10; passive responder after connect, Loop A). At the content level the projection is unrequested — there is no query parameter, filter, or state argument in OBJ-2 (index.ts:189-234), so the agent cannot ask for state, cannot ask for more, and cannot ask for less. It gets the same five fields for calling the tool
Selection signal None. The payload is a fixed five-field shape computed unconditionally at index.ts:106-117. No relevance, recency, similarity, query, or budget signal exists (ABS-9, ABS-3). Selection is degenerate — constant shape, aggregate values
Targeting None. Object.keys() and .length are whole-store aggregates. No per-entry addressing exists anywhere in the boundary; there is no retrieval primitive to target with
Selection scope and budget Scope: the entire process-global store since process start — not conversation-scoped, because thinkingServer is a module-scope singleton (index.ts:249) with no session partition. Budget: unmanaged. thoughtHistoryLength is O(1) in payload size regardless of store size; branches grows linearly with the number of distinct caller-supplied labels and is never truncated, capped, or token-budgeted. The one unbounded payload component is thus under the caller's own control
Delivery and consumption point Delivered as a single MCP text content block holding JSON.stringify(..., null, 2) (index.ts:109-115), returned from the CallToolRequestSchema handler (index.ts:255-258) over stdio. Framing over the wire is the SDK's (SRC-6, uninspected). The consumption point — whether and where the host places tool results in model context — is host-owned and excluded from the boundary. Status is affordance only: the server delivers to the transport; BAP-2's horizon is one turn "unless the host retains it"
Behavioral-faithfulness test Absent (B0). No test, fixture, assertion, trace, or run artifact exists (ABS-8), and no in-code check compares the returned counters against anything. Prevents any conclusion that delivery, parsing, or interpretation of the projection is faithful, and prevents any conclusion that the counters are correct in the presence of concurrency

RTE-2 — tool advertisement (index.ts:251-253), annotated as a negative

Field Annotation
Read-back direction Pull (host-elicited tools/list) — but what is delivered is static shipped material (OBJ-1, OBJ-2), which under the given definition is retained state, not memory read-back
Selection signal None. Unconditional return of one static array literal; the handler ignores its request argument entirely (index.ts:251-253)
Targeting None
Selection scope and budget Fixed: 54 lines of prose plus a 9-property schema, byte-identical on every call. Does not grow, shrink, or vary with use, session, or accumulated state
Delivery and consumption point tools/list response over stdio; placement into model context is host-owned and uninspected (BAP-1: "the server affords, the host disposes")
Behavioral-faithfulness test Absent (B0, ABS-8)

The load-bearing annotation: RTE-2 never consults OBJ-4 or OBJ-5 (ABS-3; runtime §5.2 Loop B records "state reads and writes: none"). There is no mechanism by which accumulated use changes what is advertised to the model — no adaptive description, no learned schema, no history-conditioned instruction. The system's one authority-bearing channel toward the model (BAP-1) is fixed at build time.

4. Four separate findings

Presence, wiring, activation, and causal effect are recorded separately with separate evidence. None may be upgraded into the next.

4.1 Context presence

Sub-claim Status Evidence Ceiling / prevented conclusion
Accumulated material (MEM-1, MEM-2) is placed onto a route directed at the receiving agent implemented index.ts:113-114 read the store; :106-117 place the values into the returned content block Cannot be raised: whether the block enters model context is host-owned and uninspected (host excluded, evidence-packet.md §2). Presence in context is therefore not established — only presence on the wire
Stored thought content is placed onto any route toward the model absent (B0) ABS-3: rg 'thoughtHistory\|branches' yields 8 hits, all enumerated in RTE-6/RTE-7/RTE-9; the only reads are Object.keys() and .length at index.ts:113-114 Prevents any claim that the server retrieves, returns, summarizes, or re-serves prior thoughts
Stored thought content reaches a human implemented RTE-8, index.ts:103-104, full thought text to stderr on every valid call Not read-back (see §2.5). Whether any human reads the stream is outside the boundary
Static shipped instruction (OBJ-1) is placed onto a route toward the model implemented RTE-2, index.ts:251-253 Static material, not memory. Presence in model context host-owned, uninspected

4.2 Deployed wiring

Sub-claim Status Evidence Prevented conclusion
The read-back route is wired end-to-end inside the inspected artifact implemented Handler registered at index.ts:255-267; processThought reached by name equality at :256; singleton constructed at module scope :249 before runServer() Dispatch into the handler is the SDK's (SRC-6, uninspected). Prevents any claim about request ordering or concurrency, and therefore about whether OBJ-4 order corresponds to call order under load
The wiring inspected is the wiring that runs in a deployment not established CMP-9/RTE-13: tsc → dist/index.js, files: ["dist"] (SRC-3:10-20); the README's configuration (SRC-2:47-59) launches npx -y @modelcontextprotocol/server-sequential-thinking with no version pin Combined with the ~20-month boundary age, this prevents any claim that a deployment following the documented configuration runs the analysed memory code
A store exists per conversation not established; the code affords the opposite Singleton at index.ts:249, no session key, no request identity, no partitioning Prevents any claim that thoughtHistoryLength counts one conversation's thoughts. If a host multiplexes conversations onto one process, they interleave in one array with no separator. Whether hosts do this is outside the boundary

4.3 Activation

Status: uninspected. Not absent — an absence would require a search boundary in which activation could have appeared, and no such boundary exists here.

Evidence for the status: the behaving agent (the host LLM) is excluded from the boundary (evidence-packet.md §2), and ABS-8 records that no run artifact, trace, or log exists anywhere in the boundary that could show a behavior change. There is no intervention, ablation, or comparison available.

Prevented conclusions: any claim that delivered material changed model behavior; any claim that the tool descriptor (OBJ-1, BAP-1) changes what the model does; any claim that the returned counters are read, parsed, or acted on.

The strongest statement available is a bound on what activation could be: the read-back channel carries only an accepted-call count and a set of branch labels (§2.4), so any behavior change attributable to memory in this system would have to be one that a count and a label list can support. It could not be a change grounded in the content of a prior thought, because that content never returns (ABS-3).

4.4 Causal effect

Status: uninspected. No intervention, no comparison, no A/B, no with/without run exists inside the boundary (ABS-8), and both halves of the loop that would have to show the effect — host and model — are excluded.

Prevented conclusions: any claim that the retained store improves, structures, lengthens, or otherwise changes reasoning; any claim that branching or revision labelling has an effect on outcomes; any claim about the difference between calling this tool and not calling it.

Doctrine in OBJ-1/OBJ-9 asserts reasoning outcomes (CLM-1–CLM-3, CLM-5, CLM-10), and that doctrine is claimed, at advisory force (BAP-1), addressed to the model. It supplies no evidence at any status about effect, and it is not upgraded here. Its truth, scope, and warrant are the epistemic lens's to record.

5. Authority references

Referenced by ID; consumer, channel, force, and horizon are read off the register rather than inferred from the family label.

Path Bearing on this lens
BAP-1 Consumer host LLM, channel RTE-2, force advisory, horizon host-determined. The material it carries is static shipped, not accumulated. So the system's only instruction-force path toward the model is fixed at build time and is not a memory path
BAP-2 Consumer host LLM, channel RTE-9, force informational only — obliges nothing, horizon one turn unless the host retains it. This is the read-back path. Its authority horizon (one turn) is shorter than the retention horizon of the material it reports on (process lifetime), which is in turn shorter than the conversation that produced it
BAP-3 Consumer human log reader, channel RTE-8 stderr, force display only, no consumer inside the boundary. The richest content channel carries no behavioral force here
BAP-4 Consumer host/model at call time, channel OBJ-2 plus in-process re-validation RTE-4, force enforcing, horizon every call for the process lifetime. This is the only gate on what enters the store — and it enforces envelope shape, not content (ABS-5). Four fields are checked; revisesThought, branchFromThought, branchId, needsMoreThoughts are cast without check (index.ts:50-54), so caller-supplied values of any type enter the store and, for branchId, become index keys
BAP-5 Consumer human operator, advisory setup. No memory bearing beyond RTE-13's unpinned launch, recorded in §4.2

No path in the register has a memory-derived consumer with binding force. The one enforcing path (BAP-4) constrains input into the store; nothing constrains, or is constrained by, what comes out of it.

Lineage and curation labels are kept independent of transformation, acceptance, and warrant throughout: the append at index.ts:94 is retention, and retention establishes nothing about semantic preservation, correctness, or acceptance of what was retained.

6. MEM-* proposals and corrections to registered records

Proposals (lens-local until the orchestrator registers them)

Tag Proposed record Concrete identity Why it needs its own ID
MEM-1 History cardinality — the accumulated scalar view of OBJ-4 index.ts:114 (this.thoughtHistory.length), written implicitly by the append at index.ts:94 It is one of exactly two accumulated parts with a consumer. OBJ-4 as registered bundles a content array with a cardinality projection whose consumer, form, and authority path all differ (BAP-2 vs no consumer at all)
MEM-2 Branch-key set — the accumulated key view of OBJ-5; monotonic, model-authored strings created index.ts:97-99, read index.ts:113 (Object.keys(this.branches)) Same split as above for OBJ-5, and it is the only accumulated caller-authored text that returns to the caller. Bundling it with bucket content would hide that
MEM-3 Normalized-record projection — the lossy 9-field whitelist that determines what actually enters the store index.ts:45-55, the object literal returned by validateThoughtData It bounds what the store could ever hold or return: keys outside the whitelist are dropped, and the retained object is a fresh copy, not the caller's argument object. This is a lineage transform distinct from the RTE-4 gate it sits inside
MEM-4 Reference aliasing between OBJ-4 and OBJ-5 — branch buckets hold the same object references as history entries index.ts:94 (push(validatedInput)) and index.ts:100 (push(validatedInput)) Determines that OBJ-5 is an index over OBJ-4, not a second copy — the correct reading of the "stored twice" fact. Recorded in the runtime account's Loop C prose but carries no ID

Two facts are cited against existing IDs rather than proposed as new records, since they are already covered: the pre-store totalThoughts overwrite is RTE-5 (index.ts:90-92), and its memory-lens consequence is that the retained record diverges from the submitted value in one field with no trace of the original; the silent branch-index miss on falsy branchId/branchFromThought is a property of RTE-7 (index.ts:96).

Corrections to registered records

None. Every registered record this lens depended on (OBJ-3–OBJ-8, RTE-2, RTE-5–RTE-10, ABS-2–ABS-9, BAP-1–BAP-4) was checked against index.ts and found accurate, including ABS-3's claim of exactly two read sites and ABS-4's claim that needsMoreThoughts has no consumer (verified: declared :21, copied :54, documented :175, schema'd :228, read nowhere). All contributions above are extensions by ID, not re-inventories.

One scoping note, not a correction: ABS-3 is scoped to stored thought content, and this lens relies on that scoping. MEM-2 shows that accumulated caller-authored identifiers do return via Object.keys at index.ts:113. Reading ABS-3 as "nothing accumulated ever returns" would be wrong; as written it is correct.

7. Limitations, each with the conclusion it prevents

Limitation Conclusion it prevents
ABS-8 — no test, fixture, trace, log, or run artifact anywhere in the boundary Prevents every observed and causally supported status in this lens. Caps §4.1 and §4.2 at implemented and leaves §4.3 and §4.4 at uninspected
The host is excluded from the boundary (evidence-packet.md §2) Prevents any claim that the read-back projection or the tool descriptor enters model context. Context presence is established on the wire only
The host LLM is excluded from the boundary Prevents any claim of activation or causal effect, and any claim about what a model does with the counters
SRC-6 — @modelcontextprotocol/sdk 0.5.0 uninspected, node_modules/ absent Prevents any claim about wire framing of OBJ-6, about pre-handler validation, and about request ordering or concurrency — and therefore prevents any claim that OBJ-4's order corresponds to call order, or that thoughtHistoryLength is race-free
Host stderr handling is uninspected Prevents the stronger reading of ABS-2. ABS-2 licenses "the server performs no persistence"; it does not license "thought content is never durably retained anywhere", because RTE-8 writes full thought text to a host-owned sink
Boundary age (~20 months) plus the README's unpinned npx -y launch (RTE-13, SRC-2:47-59) Prevents any claim that a deployment following the documented configuration runs the memory behavior analysed here, and any claim about the current upstream state of this server
Multiplexing behavior of hosts is outside the boundary Prevents any claim that the store is conversation-scoped, and therefore prevents interpreting thoughtHistoryLength as a count of one conversation's thoughts. The code affords sharing; whether hosts share is not established