MemPalace synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From host activity to retained memory
MemPalace is a memory and context-supply system with shipped host adapters and a delegated-task coordination plane. It does not own the enclosing agent's answer generation. This account covers its repository at commit 1392e1810c3ca67776bd7fe3e198788a96dbd6d1, with an applicability cutoff of 2026-10-14. The whole-system boundary includes capture, maintenance, shipped instructions, later consumers and evaluators; it does not imply exhaustive implementation inspection. SRC-1 supplies code and doctrine/design. No host session, live palace, executed test, benchmark rerun or causal experiment was supplied or produced. The evidence establishes implementation paths and attributed intent, not their observed exercise or benefit. See boundary and evidence and Source register.
The ordinary traced path starts with an initialized Hermes host session. A completed-turn callback supplies user and assistant text to a bounded queue. A background worker classifies its filing location and upserts a drawer: retained text organized by wing and room. Software admits each exchange without per-content human approval. The host must invoke the lifecycle callbacks; backend initialization can leave the adapter inactive. Queue saturation drops turns, filing errors are logged rather than durably retried, and shutdown waits only five seconds. This is best-effort capture, not a confirmed lossless save. RT-RTE-hermes-capture.
CLI imports and transcript hooks are alternatives, not the same capture guarantee. Hooks schedule a daemon job or detached CLI process; submission does not confirm final persistence. Once daemon submission is attempted, its error branch does not replay locally because double-mining is possible. Downstream filing normalizes and chunks source content. General extraction rejects segments by marker scores and thresholds; normalization can spell-correct user text. Message sweep has its own cursor and IDs and can overlap file mining. Thus exact retention of an admitted representation must not be confused with byte-preservation of every original input. RT-RTE-hook-ingest, MEM-RTE-source-filing.
Several retained parts perform different jobs. Drawer words provide recallable content; symbolic metadata filters and reconstructs chunks; dense embeddings influence candidate selection. Separate closet lines are reduced topic/quote/summary locators pointing to drawers, not complete transcripts. Optional LLM regeneration groups retained drawers by source and automatically admits parsed locator output after an operator starts the job. Its input and packing caps omit material; parsing does not verify quote fidelity or factual accuracy. Embedding and generation clients do not establish an in-use weight-training step, while exact model versions and provider-internal changes remain branch-specific uncertainties. MEM-OBJ-drawer-text, MEM-OBJ-drawer-metadata, MEM-OBJ-drawer-vectors, MEM-OBJ-closet-lines, MEM-RTE-closet-regeneration, runtime components.
Admission, revision and later reliance
Admission is not one universal approval gate. Interactive setup lets humans edit people/projects; --yes bypasses review, and topics pass through automatically. Explicit tools structurally validate caller content, but caller identity alone does not establish human authorship or approval. MCP dispatch can veto operations by read-only and runtime-state checks. This is a server-owned request boundary: direct CLI/API/callback writes do not inherit it, and permitted recovery can touch disk. Hub forwarding refuses local replay of classified mutations after eligible remote failures; it does not guarantee exactly-once execution. RT-RTE-init-detection, RT-RTE-mcp-dispatch, RT-CLM-readonly-boundary, RT-RTE-hub-fallback.
Checkpoint saving adds a consequential similarity check: nearest-hit comparisons can mark an item duplicate, and the checkpoint consumer then skips it. Dedup errors instead allow filing. Embedding closeness is not semantic equivalence; a near match may contain a correction. Explicit update/delete provides a separate revision route and purges affected source closets after content changes. Neither storage visibility checks nor replacement certify the proposition's truth. EPI-RTE-duplicate-check, EPI-RTE-checkpoint-disposition, EPI-RTE-drawer-revision.
Temporal facts have a distinct lifecycle. SQLite relationship rows carry validity intervals and optional provenance. Caller-requested invalidation ends eligibility while preserving history; supersession transactionally closes old rows and inserts or reuses a successor. A reused open successor does not have its start aligned to the requested boundary. If it starts later, an as-of query at that boundary can exclude both old and successor rows despite successful supersession. This is an implementation deduction, not an observed failure. Temporal grouping and default confidence do not independently endorse a relationship. MEM-OBJ-kg-triples, MEM-RTE-temporal-facts.
Selection into a later invocation
Requested search is a pull route: query embeddings, filters and lexical/closet scoring select retained material for MCP text or a host context return. Closet-assisted hydration can choose a best passage plus neighboring chunks and truncate it rather than return the whole logical drawer. Direct logical drawer GET is the fuller read alternative. A search miss therefore does not prove stored absence, and a hit need not contain every relevant qualification. The external host/agent owns interpretation and answer generation. RT-RTE-palace-recall, memory read-back annotation, RT-BAP-agent-context.
Independent startup supply is different. DSH registers agent-local prompt providers; Hermes exposes identity and cached wake-up material. The startup selector reads full identity text and selects weighted/recent snippets within a candidate window of up to 2,000 filings: at most 15 drawers, each capped at 200 characters, with an approximately 3,200-character L1 content budget. The identity file is not token-capped; headings and backend scan limits prevent a universal total-token or globally highest-importance guarantee. DSH's first assembly waits only within its time budget. Disposal removes registrations, not content already consumed elsewhere. Host compaction and child-agent context remain host-owned; Hermes's pre-compression callback is a no-op. MEM-RTE-wake-context, RT-RTE-palace-recall.
For session continuity, silent Python hooks retain short prefixes from recent eligible user messages, not assistant reasoning or a complete session summary. Later diary reads select recent physical rows, which may be chunks. Prompted summaries and imported daily files have different, partly unresolved authorship and completion controls. DSH does not relay the non-silent hook's request for model-authored saving, so checkpoint completion is not equivalent across hosts. MEM-RTE-diary-checkpoint, report verification limits.
Media search returns descriptors, paths and availability, not automatically original bytes. Tunnels support explicit navigation and persist traversal counters, but counter display does not establish default semantic-ranking influence. Explicit maintenance changes future stored eligibility without proving synchronous withdrawal from existing host prompts, backups or replicas. These differences matter when the scenario requires payload consumption, learned navigation or complete forgetting rather than reference delivery. MEM-RTE-media-index, MEM-RTE-tunnel-traversal, MEM-RTE-memory-maintenance.
Delegated work and its delivery boundary
The representative multi-step workflow is a coding task through the logstream. Shipped instructions connect normalized task approval, creation, handoff, worker claim, patch return and requester review. Approval of the exact task is doctrine, not enforced by the constructor. The portable default is a pasteable pointer requiring the worker to reload the persisted goal, definition of done and delivery instructions. An event alone does not wake a non-monitoring agent. Acknowledgments are linked append-only reports, not exclusive execution locks. WF-RTE-task-create, WF-RTE-event-ack, RT-BAP-task-handoff.
Optional controlled launch checks addressed identity, branch, resolved base commit and clean workspace before starting Codex or Claude. A mismatch rejects launch; the instructed recovery is a corrected superseding request, not automatic workspace repair. The external runner owns execution and permissions. Patch submission stores an artifact and ready event, even allowing format warnings; it does not apply the patch. Requester hash checking, explicit application, task-specific verification and applied/failed acknowledgment are downstream policy. Filing a searchable outcome drawer is prescribed, not automatic. No successful instance or acceptance trace was observed. RT-RTE-task-launch, RT-RTE-patch-delivery, WF-CLM-review-closure, workflow account.
Assessment and evidence that would change it
The strongest supported learning-related contribution is automatic trace-fed production of durable continuity and retrieval artifacts. Session-prefix checkpoints and session-fed extracted segments/closet locators have later context or ranking consumers. This is wired trace learning, distinct from raw-log retention and from demonstrated improved capacity for future action. The unresolved question is whether these transformations make later work better rather than merely different. No before/after capacity comparison or component intervention was supplied. MEM-RTE-diary-checkpoint, MEM-RTE-source-filing, MEM-RTE-closet-regeneration, profile verification.
Other properties remain separate. For the inspected capture, setup and task paths, a formulated-theory loop meeting conditions 1–4 is not established: localized operational guidance is inspectable, but theory-dependent decisions, content-directed criticism with blame and revision, and retention shaping another criticism round are not demonstrated as a connected theory route. This is not a whole-system absence claim. Criticism of a consumed theory improving future capacity is independently unestablished. Traversal self-recording and organization auditing support narrower self-state mechanisms, not a demonstrated reflective theory builder. Autonomous queueing and selection are implemented, but a fully autonomous theory builder is not established, and workflow review remains assigned to a requester. RT-RTE-hermes-capture, RT-RTE-init-detection, MEM-RTE-tunnel-traversal, EPI-RTE-palace-audit, epistemic bounded conclusion.
Exercised self-improvement at this boundary is likewise unestablished. Derived memory can enter later behavioral context, but no inspected comparison closes objective-bearing evidence shaping an update, change to MemPalace's own organization, live consequential consumption, and subsequent operation causally dependent on that change. Delivered coding patches remain proposals until externally applied. This does not negate the implemented memory updates or require weight training. RT-RTE-patch-delivery, WF-CLM-review-closure, MEM-RTE-closet-regeneration.
For prior-work lookup, the architecture supplies persistent material and several reload routes. For guaranteed exact original recall or evidence-backed truth acceptance, the inspected mechanisms do not establish the required guarantee. Narrow formula-derived relevance tags, audit scores and retrieval metrics do not provide a universal answer oracle. README's raw 96.6%, held-out hybrid 98.4% and full-set reranked ≥99% R@5 figures are attributed, unreproduced retrieval claims under different conditions, not QA accuracy or live-loop validation. NDCG here uses an ideal ordering drawn only from relevance labels already in returned top-k, preventing an unqualified standard-NDCG comparison. RT-CLM-recall-protocol, EPI-RTE-relevance-tags, EPI-RTE-palace-audit, EPI-CLM-benchmark-recall, EPI-RTE-retrieval-evaluation, report verification limits.
Paired source-to-capture-to-delivery host traces would change the activation and fidelity assessment; comparable later-task outcomes with independent component variation would change benefit attribution. Evidence-consuming content acceptance with a named use and scope would change the warrant assessment. An installed, subsequently exercised change tied causally to objective-bearing evidence would change self-improvement attribution. A temporal implementation that aligns or rejects incompatible reused successors would remove the specific replacement-boundary counterexample. These are different evidence or system changes, not one missing universal test.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operational or causal evidence | SRC-1; RT-RTE-hermes-capture, RT-BAP-agent-context, EPI-RTE-retrieval-evaluation | Frozen implementation/doctrine; no live palace, host trace, executed test or rerun | Observed persistence, compliance, reliability, latency, answer benefit and component causation | Inspectable capture/read/answer traces and comparable outcome experiments at named configurations |
| Capture, transformation and delivery are not end-to-end verbatim guarantees | SRC-1; RT-RTE-hermes-capture, MEM-RTE-source-filing, MEM-RTE-diary-checkpoint, RT-RTE-palace-recall | Queue, normalization/extraction, prefixes and selected hydration | Lossless original-input-to-answer recall or absence inferred from no hit | Paired raw input and retained/delivered artifacts; stronger capture acknowledgments and explicitly fidelity-preserving paths |
| Host checkpoint inventory is incomplete; DSH does not relay non-silent model-save prompts | SRC-1; MEM-RTE-diary-checkpoint, RT-RTE-hook-ingest, report verification limits | Python hook paths plus verifier inspection of DSH autosave | Equivalent prompted-checkpoint completion across hosts | Branch-specific host traces or inspected relay/confirmation implementations; preserve silent-prefix and transcript-submission positives |
| Graph normalization and contradiction-check routes lack separate analyst tracing | SRC-1; MEM-RTE-temporal-facts, MEM-OBJ-kg-triples, epistemic coverage, report verification limits | Verifier inspected retained model rewrite proposals, explicit apply and KG-relative checker exit; analyst inventory remains partial | Complete graph-curation, model-authorship and evaluator-consumer comparison; system-complete absence of epistemic routes | Separate records tracing proposal/review/temporal rewrite and checker consumers, including semantics and downstream use |
| NDCG denominator differs from a full-reference ideal ranking | SRC-1; EPI-OBJ-retrieval-metrics, EPI-RTE-retrieval-evaluation, report verification limits | LongMemEval formula uses labels present in returned top-k | Comparison with NDCG normalized against all available relevant items | Metrics recomputed against a specified full reference set; inspected dataset and run artifacts |
| Original media bytes lack their own storage unit in the partial profile inventory | SRC-1; MEM-OBJ-media-asset, MEM-RTE-media-index, profile verification limits | Indexed references/descriptors/vectors plus original bytes in external local files | Complete object-by-object storage comparison or inference that descriptor/index units enumerate byte retention | Explicit original-file storage unit and retention/consumer tracing; existing files union remains supported |
| Reused temporal successor may not cover the replacement instant | SRC-1; MEM-RTE-temporal-facts | Inspected supersede reuse branch and actual as-of bounds | Unconditional immediate successor visibility after successful supersession | Interval-consistent insertion/reuse policy or rejection; branch-specific validation evidence |
| Source truth, initial authorship and provider behavior remain unresolved | SRC-1; MEM-OBJ-drawer-text, MEM-OBJ-identity-text, MEM-CMP-closet-provider, RT-CMP-openai-compat | Caller/import content and inspected provider clients, not external implementations | Independent factual acceptance, trusted authorship, stable installed model identity or provider-retention guarantees | Content provenance and evidence-consuming acceptance; pinned model/provider artifacts and retention contracts |
| Backend, parser, host, graph and recovery alternatives remain incompletely inspected | SRC-1; MEM-OBJ-drawer-vectors, RT-RTE-backend-registration, MEM-RTE-tunnel-traversal, MEM-RTE-memory-maintenance, boundary coverage | Named native/remote stores, formats, legacy wrappers, graph policy, backups/migration/replicas and evaluator modes | Universal isolation, fidelity, graph adaptation, restore or global erasure; complete mechanism enumeration | Focused traces and contracts for each material alternative, including cached/replicated withdrawal |
| Coordination delivery is not verified product change | SRC-1; RT-RTE-task-launch, RT-RTE-patch-delivery, WF-RTE-event-ack, WF-CLM-review-closure | Starting-state gates, appended artifacts/statuses and review doctrine | Exclusive execution, task completion, installed correct patch or exercised self-improvement | Worker reload/execution, requester application/test evidence, and causally linked subsequent operation |
No unresolved cross-report conflict was declared in the supplied reconciliation. The table carries all three report-verification limits and the profile-verification limit without converting the remaining scoped uncertainty into absence. Source reads for this synthesis: none.