Instinctual Memory agentic-system analysis
Type: types/agentic-system-analysis-overview.md
Boundary and evidence
This analysis evaluates how Instinctual Memory captures, curates, stores, retrieves, changes, and injects durable memory for AI-agent use, including its trust and integration mechanisms. The target is the repository's complete mem system at commit 6acb13dc35765bf5ccfc87e445dd09c480f1c28a; the repository's module and source inventory locates the Rust CLI and library, memory journal and Git-backed fact store, ingestion and consolidation paths, search and reranking, validated changes and erasure, hooks and setup, MCP and HTTP interfaces, shell, and evaluation facilities within scope. The README and /mem skill provide declared operating intent and integration guidance; sampled implementation inspection covers the crate/module surface, CLI entrypoint, and MCP interface, while later jobs can inspect additional paths within this frozen source.
The whole-system boundary includes the repository's code and bundled skill as shipped. It treats Git as the store's persistence and publication substrate; configured local or remote model services and local reranker models as external dependencies; and Claude Code, Codex, OpenCode, other MCP clients, project files, and host session exports as integration participants. The host agent's orchestration, model behavior, conversation selection, and adherence to injected memory are excluded: their contribution to end-to-end agent behavior cannot be established from this repository. External model behavior and service retention are also excluded, preventing conclusions about provider-side handling or the correctness of model-generated extraction/reranking. User environments and actual deployed runs are not supplied, preventing claims about installation success, observed use, or production outcomes. The frozen evidence is the repository at the reviewed commit; the analysis cutoff is 2026-09-30. Overall evidence tier: code-grounded.
Source register
| source ID | kind | identity/location | revision or capture | evidence layer | inspected scope | citation anchors | access gaps and conclusion prevented |
|---|---|---|---|---|---|---|---|
| SRC-1 | git | https://github.com/jasonkneen/instinctual-memory; operational access root /home/zby/llm/commonplace/related-systems/jasonkneen--instinctual-memory |
6acb13dc35765bf5ccfc87e445dd09c480f1c28a |
implementation and doctrine/design, in separately identified scopes | Implementation: crate manifest, library/module surface, CLI entrypoint, MCP implementation. Doctrine/design: README and bundled /mem skill. |
README.md, Cargo.toml, src/lib.rs, src/main.rs, src/mcp.rs, skills/mem/SKILL.md |
No deployed host, configured store, external model service, or observed agent session was inspected; this prevents conclusions about live operation, provider behavior, and whether consuming agents use or benefit from retrieved memory. |
Reconciliation
The members converge on the code-grounded progression from source-event retention to candidate fact extraction or caller-supplied changes, validated Git publication, and later retrieval or writeback (OBJ-1, OBJ-2, OBJ-3, MEM-OBJ-1, MEM-OBJ-2, MEM-OBJ-3, RTE-1, RTE-4, RTE-5, RTE-6, MEM-RTE-1, MEM-RTE-3, MEM-RTE-4). Runtime and epistemic accounts independently characterize the extraction evidence check as traceability and structural/domain admissibility, not a factual truth test (RTE-4, EPI-RTE-1). This is convergence on inspected code, not independent evidence of operation. No member supplies an observed candidate, live host operation, or causal comparison (RTE-1, RTE-2, RTE-3, RTE-4, RTE-5, RTE-6, RTE-7, MEM-RTE-1, MEM-RTE-2, MEM-RTE-3, MEM-RTE-4, EPI-OBJ-1, EPI-OBJ-2, EPI-OBJ-3, EPI-OBJ-4).
No established fact requires amendment. The runtime routes remain the canonical execution and integration account. The memory routes retain their memory-specific grouping and admission details. The epistemic routes characterize checks, transformations and authority functions without establishing separate dispatch paths.
Amendment: MEM-RTE-2 is superseded by RTE-1; both identify Claude session-start and prompt hooks selecting retained facts and delivering them through additionalContext (SRC-1 src/hook.rs, src/setup.rs; RTE-1, MEM-RTE-2). The memory-specific selector findings remain in the memory member and profile.
Amendment: MEM-RTE-4 is superseded by RTE-6; both identify operator-requested selection of eligible facts and writing a managed block into a host instruction file (SRC-1 src/cli.rs; RTE-6, MEM-RTE-4). The affected writeback route identity is represented by RTE-6; the memory-specific selection and pull findings remain in the profile and memory report.
MEM-RTE-1 groups ingestion, backfill and consolidation. It overlaps the hook synchronization in RTE-1 and the direct ingestion/consolidation route in RTE-4, but its scope is not identical to either runtime record. MEM-RTE-3 groups requested retrieval and change operations over several surfaces; RTE-2, RTE-3 and RTE-5 distinguish those surfaces and operations. Both memory records remain as lens-specific groupings, not duplicate route identities.
EPI-RTE-1 characterizes the evidence-check function within RTE-4; EPI-RTE-2 characterizes JEV relevance scoring within search on RTE-2 and RTE-3; EPI-RTE-3 characterizes eligibility filtering used on RTE-1, RTE-2 and RTE-3. They retain epistemic function and target detail while the runtime records own the broader execution routes. EPI-OBJ-1 and EPI-OBJ-2 distinguish extracted and caller-supplied subsets of OBJ-2 by provenance; EPI-OBJ-4 identifies the host-file rendering related to OBJ-3. These are scoped content or form distinctions, not identical objects to supersede. EPI-OBJ-3 is ephemeral relevance ordering, not truth-apt retained content. No epistemic record establishes truth acceptance or host uptake.
The memory-comparison profile has been checked axis by axis against the memory, runtime and epistemic records, including each epistemic transformation object. storage_substrate values files and repo match the journal and Git tree/access structures (MEM-OBJ-1, MEM-OBJ-2, MEM-OBJ-3, OBJ-1, OBJ-2). representational_form values natural-language and symbolic match the event text, fact statements and structured fields; EPI-OBJ-1 and EPI-OBJ-2 retain text propositions with structured fact fields, while EPI-OBJ-4 renders stored statements as natural language (MEM-OBJ-1, MEM-OBJ-2, MEM-OBJ-3, EPI-OBJ-1, EPI-OBJ-2, EPI-OBJ-4). lineage values authored, imported and trace-extracted are supported by explicit change requests, file ingestion and trace-fed consolidation (MEM-RTE-1, MEM-RTE-3, EPI-OBJ-1, EPI-OBJ-2, RTE-4, RTE-5).
The behavioral_authority values instruction, knowledge and ranking describe, respectively, writeback into a host instruction channel, facts made available through hooks or requested reads, and relevance ordering of search results; none establishes enforced host reliance (RTE-1, RTE-2, RTE-3, RTE-6, BAP-1, BAP-2, BAP-3, EPI-OBJ-3, EPI-OBJ-4). write_agency automatic and manual is supported by thresholded hook backfill/consolidation and operator-invoked ingestion, changes and writeback (MEM-RTE-1, MEM-RTE-3, MEM-RTE-4, RTE-1, RTE-4, RTE-5, RTE-6). The corrected curation_operations values evolve and invalidate match correction/supersession, source refresh and forget; candidate rejection is not classified as dedup (MEM-RTE-1, MEM-RTE-3, RTE-4, RTE-5). No amendment is needed for the specialist's explicit withdrawal of the former dedup mapping.
read_back_direction pull and push and read_back_signal identifier and inferred-lexical match requested reads/writeback and the session-start ID lookup plus prompt lexical selector (MEM-RTE-2, MEM-RTE-3, MEM-RTE-4, RTE-1, RTE-2, RTE-3). trace_learning "yes" is supported only as wired automatic trace-fed durable memory that can reach later consumers; it does not claim observed learning or improved task performance (MEM-RTE-1, MEM-OBJ-1, MEM-OBJ-2, EPI-OBJ-1, RTE-4). trace_source partial appropriately leaves the unclassified adapters open (MEM-RTE-1, RTE-4). learning_scope distinguishes the claimed cross-task target from wired per-project scoping; learning_timing staged matches journal backfill followed by thresholded consolidation (MEM-RTE-1, MEM-RTE-2, RTE-1, RTE-4). distilled_form natural-language and symbolic matches fact statements and their typed fields (MEM-OBJ-2, EPI-OBJ-1, EPI-OBJ-2). faithfulness_tested "no" is consistent with the retrieval-rubric claim: the evaluation scores search output and does not test whether task behavior depends on recalled content (MEM-CLM-1, EPI-OBJ-3). The partial coverage and claimed-versus-wired qualifications remain in the profile; no value has been strengthened.
Bounded synthesis
Evidence basis: the whole-repository boundary at Instinctual Memory commit 6acb13dc35765bf5ccfc87e445dd09c480f1c28a, with implementation and bundled design inspected but no configured deployment or observed host run (SRC-1).
Instinctual Memory appends source events to a journal, then rules or a configured language model can propose durable facts for a Git-published tree. The extractor can run after session-end backfill or through direct operator commands; caller-requested remember and correction operations provide another write path (OBJ-1, OBJ-2, RTE-1, RTE-4, RTE-5). The extraction check ties evidence text to a source event and applies schema/domain conditions, but no truth oracle or per-fact human approval is established. The semantic relation between an extracted statement and its source remains indeterminate; persistence does not amount to epistemic acceptance (EPI-RTE-1, EPI-OBJ-1, EPI-OBJ-2). Normal publication uses validated compare-and-swap; that protocol constrains writes, not factual warrant (CLM-2, RTE-4, RTE-5).
Later access comes through automatic Claude hooks, requested reads/searches, or operator-requested writeback (RTE-1, RTE-2, RTE-3, RTE-6). The start hook selects a preference by identifier; prompt retrieval uses lexical overlap and a six-fact limit. Explicit search can use relevance reranking. Eligibility, selection and ranking affect which facts are presented and in what order, not their truth status (EPI-RTE-2, EPI-RTE-3). Writeback places selected facts in a host instruction-file block. These mechanisms wire context delivery or file writing, but host loading, model uptake and behavior change were not observed (BAP-1, BAP-2, BAP-3, EPI-OBJ-4).
The runtime pass does not establish a theory route meeting conditions 1–4: it does not establish a consumed localized theory, content-directed criticism, or iteration using criticism (RTE-4). Whether criticism improves capacity for future action is uninspected. The system is not established as reflective or autonomous at this boundary, and its wired trace-fed memory path does not establish self-improvement (RTE-1, RTE-2, RTE-3, RTE-4, RTE-5, RTE-6, RTE-7, MEM-RTE-1, EPI-OBJ-1). There is no task-level faithfulness or causal benefit evidence (MEM-CLM-1).
Limitations
| limitation | affected source, record, or route IDs | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| No installed host, configured store, live operation, or model-service response was inspected. | SRC-1, RTE-1, RTE-2, RTE-3, RTE-4, RTE-5, RTE-6, RTE-7, MEM-RTE-1, MEM-RTE-2, MEM-RTE-3, MEM-RTE-4, EPI-OBJ-1, EPI-OBJ-2, EPI-OBJ-3, EPI-OBJ-4 | Repository commit 6acb13dc35765bf5ccfc87e445dd09c480f1c28a; external host/provider operation excluded. | Actual invocation, output, candidate state, host activation and successful deployment are not established. | A retained execution trace or artifact from a configured host and store, with provider interactions identified where relevant. |
| No evidence-consuming truth criterion or independent factual reviewer was identified in inspected paths. | RTE-4, RTE-5, EPI-RTE-1, EPI-OBJ-1, EPI-OBJ-2 | Same frozen repository boundary; extraction, change, validation and publication paths inspected. | No produced or caller-supplied fact is established as true or supported by its cited evidence. | A specified truth/evidence criterion and retained candidate-level results showing its application. |
| The semantic relation between extracted statements and source events is not determined by text-occurrence checks. | RTE-4, EPI-RTE-1, EPI-OBJ-1 | Consolidation/check implementation at the frozen commit; no candidate instance supplied. | Whether extraction preserves, entails, reshapes or adds propositions, and any candidate-linked discovery lifecycle, remain unresolved. | An observed candidate and source event with semantic comparison, followed by evidence for any claimed criticism or revision stages. |
| Host use and behavioral effects were outside the evidence set; the shipped retrieval evaluation scores search results. | BAP-1, BAP-2, BAP-3, MEM-CLM-1, EPI-OBJ-3, EPI-OBJ-4 | Hook, operation, writeback and evaluation paths at the frozen commit; no host session or task comparison. | Faithfulness to recalled content, downstream benefit and behavioral authority in host practice are not established. | Retained task executions comparing behavior with and without returned memory, including task outcome and exposure evidence. |
| The full set of supported ingestion adapters was not classified as trace-source categories. | MEM-RTE-1, RTE-4 | Specialist traced session-log paths and a broader but incompletely mapped adapter set within the same repository boundary. | Completeness of trace_source beyond the supported session-logs value cannot be concluded. | Adapter-by-adapter mapping of retained inputs to the controlled trace-source categories. |
Verification and blockers
Semantic verification
Checked the assembled set at the whole-repository boundary, Instinctual Memory commit 6acb13dc35765bf5ccfc87e445dd09c480f1c28a, with code-grounded evidence. The overview, runtime member, byte-for-byte memory member, epistemic member, and set check agree on identity, target class, boundary, cutoff, exclusions, and the lack of deployed operation or causal evidence. The source register identifies the canonical repository and inspected scopes; all 26 distinct cited source paths exist at the frozen checkout. Checked all 19 retained quote attributions against their cited paths, commit, and ranges; each passage occurs in its range. The structural set check reports none.
Canonical IDs and ownership agree across members: SRC-1; runtime CMP-1–CMP-3, OBJ-1–OBJ-3, RTE-1–RTE-7, CLM-1–CLM-2, and BAP-1–BAP-3; memory MEM-OBJ-1–MEM-OBJ-3, MEM-RTE-1–MEM-RTE-4, and MEM-CLM-1; epistemic EPI-OBJ-1–EPI-OBJ-4 and EPI-RTE-1–EPI-RTE-3. Cross-member references resolve to the declared records. The source-dependent claims and limitations cite SRC-1 with paths; no evidenced-absence record is declared. The bounded negative findings are limited to inspected paths and the supplied evidence boundary.
The runtime routes distinguish hook delivery, MCP operations, direct CLI/shell/HTTP, ingestion/consolidation, explicit changes/erasure, writeback, and setup. Their wired implementation conclusions are kept separate from uninspected operation and host uptake. CLM-1 is scoped to hook fail-open handling; CLM-2 to normal validated compare-and-swap publication. The epistemic routes correctly characterize text matching and domain checks as traceability/admissibility, JEV as relevance ordering, eligibility as retrieval filtering, and publication or writeback as no truth warrant. Candidate transformation and correction relations remain explicitly indeterminate where code does not resolve them; no observed candidate, acceptance, theory-builder path, reflection, autonomy, learning improvement, or causal benefit is inferred.
The memory comparison has all fourteen axes, with profile records declared in the memory member and values scoped to its stated retained-memory boundary. Its classifications are supported as follows: storage_substrate files/repo; representational_form natural-language/symbolic; lineage authored/imported/trace-extracted; behavioral_authority instruction/knowledge/ranking; write_agency automatic/manual; curation_operations evolve/invalidate; read_back_direction pull/push; read_back_signal identifier/inferred-lexical; trace_learning yes; trace_source partial, session-logs; learning_scope cross-task/per-project; learning_timing staged; distilled_form natural-language/symbolic; faithfulness_tested no. The profile's claimed cross-task scope is distinct from its wired per-project scope; trace-source coverage remains partial because the other supported adapters were not classified. The yes value for trace-fed memory is expressly not a claim of observed learning or improved performance.
Reconciliation dispositions match the records: MEM-RTE-2 is superseded by RTE-1, and MEM-RTE-4 by RTE-6; the memory-specific selector, grouping, and profile findings remain available in the memory member. MEM-RTE-1 and MEM-RTE-3 remain distinct lens groupings rather than asserted duplicates. Epistemic route/object records retain their stated narrower check, provenance, and content-form scopes; they do not claim separate runtime dispatch or duplicate truth authority. No amendment changes evidence or status. The listed limitations—including absent configured host/provider evidence, absent truth criterion, unresolved extraction semantics, untested host effects, and incomplete adapter classification—are carried into the synthesis and do not prevent integration.
Deterministic validation
commonplace-validate kb/reports/state/agentic-system-analysis/AAS-2026-09-30-instinctual-memory-02/output --full runs after the manifest is written, over every member, the manifest and the set's cross-member checks; code publishes only when it passes, and publication runs it again.
Blockers
none