MentisDB synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A memory service, not the task agent
MentisDB supplies durable memory admission, selective retrieval and a versioned skill registry. Its host agent still owns task execution, tool-call timing, context eviction and interpretation of returned material. This account covers the shipped Rust engine, daemon, CLI, service transports, dashboard, Python consumer, instructions and retrieval evaluators at commit 4c9b95041da68c5a26dece9f8776ff54b05b910c, with an applicability cutoff of 2026-10-14. SRC-1 supplies inspected implementation and doctrine/design; no target execution, benchmark result or causal experiment was established. The whole-system functional boundary is not an exhaustive correctness audit. External hosts, dependencies, provider internals and deployed state are not frozen evidence. See the boundary and Source register.
The ordinary progression is caller-authored text → structural admission → retained chain → later requested selection → host context. REST and MCP serve this progression; direct Rust callers, local stdio and dashboard paths have different enforcement envelopes. The embedded operating skill asks hosts to load prior state, search before writing and checkpoint before losing context. These are available instructions, not append prerequisites. RT-RTE-memory-roundtrip, RT-RTE-access-control, RT-BAP-core-skill.
Admission retains statements without certifying them
A thought separates natural-language content from a symbolic envelope containing identifiers, attribution, types, relations and hashes. The core validates references, constructs a hash-linked record and appends through storage before updating derived retrieval state. Hash verification on reopen checks representation integrity under the storage contract; it does not establish truth, authenticated authorship or protection against rewriting an entire chain. Canonical admission precedes fallible derived work, so an error can follow a retained write. Buffered flushing and alternate adapters do not establish unconditional power-loss durability. RT-OBJ-thought, MEM-OBJ-thought-content, RT-RTE-memory-roundtrip.
Optional extraction asks an external OpenAI-compatible model for candidate statements. The parser checks JSON, recognized types and nonempty content, then trims text and normalizes metadata. It returns candidates without appending them. Caller review is declared, but no universal factual review gate or actual review episode is established. Without an input/output instance, the semantic relation between source and generated statement remains indeterminate: conservative prompting alone establishes neither faithful compression, checked derivation nor ampliative conjecture. Ordinary append similarly consumes no world evidence for the truth of supplied statements. RT-RTE-memory-extraction, EPI-RTE-extraction-schema, EPI-OBJ-extracted-thoughts, RT-CLM-extraction-review.
Maintenance changes availability and retrieval
Explicit Corrects, Invalidates and Supersedes relations suppress targeted thoughts in covered default readers while preserving their history. Optional token-set deduplication can create supersession automatically. That changes selection, not the epistemic status of either proposition: lexical similarity does not establish equivalent meaning or a truthful replacement. Audit and direct-access alternatives can expose old text. Markdown export/import, agent copying and merge can lose graph lineage and reintroduce old content under fresh identities. Whole-chain deletion also makes “permanent” history conditional on administrative retention. MEM-RTE-correction-update, MEM-RTE-markdown-exchange, MEM-RTE-agent-copy, RT-RTE-state-maintenance, MEM-RTE-chain-withdrawal.
The strongest supported trace-fed contribution is automatic construction of durable retrieval machinery. Retained conversational messages can feed embedding generation and semantic-neighbor updates; later search consumes those numerical and symbolic derivatives to select content. This is a wired trace-fed behavior-shaping update, not merely raw-log storage. It supports a local trace-learning classification without establishing a semantic lesson, trained model weights or improved capacity for future action. Actual mixed input provenance, complete backend freshness and semantic-update alternatives remain unresolved. MEM-OBJ-embedding-vector, MEM-OBJ-retrieval-state, MEM-RTE-derived-update, MEM-RTE-retrieval-evaluation, profile verification.
Inference weights are separate from these accumulated outputs. The local MiniLM adapter performs inference, with a hash-based fallback if initialization fails; its model label does not pin an exact weight digest. The extraction endpoint likewise does not freeze provider parameters. Neither adapter establishes weight training. RT-CMP-fast-embed-provider, RT-CMP-extraction-model.
Later context is selected, not restored wholesale
Recent-context returns a bounded chronological slice of valid thoughts; ranked search uses lexical, vector, graph and metadata signals and can traverse ancestors. Context bundles instead use lexical seeds and optional graph expansion to return original text with paths. A path explains connectivity, not evidential warrant. These are requested returns, not autonomous startup injection, and their limits are record/page/graph budgets rather than guaranteed model token budgets. Webhooks push new append events through a bounded, droppable queue; they do not establish historical memory read-back or an external consumer's later behavior. RT-RTE-memory-roundtrip, MEM-RTE-bundle-read, MEM-RTE-webhook-delivery.
The shipped Python adapter closes a producer-to-prompt path: it stores raw messages and later formats up to twenty type-filtered hits into chat_history. It does not target the current input, preserve original speaker class or use its configured session ID to isolate these writes and reads. Retrieval exceptions become empty history, and clear does nothing. Thus this path demonstrates prompt delivery wiring, not faithful transcript resurrection or reliable session partitioning. Memory-scope tags and branch names also do not establish identity isolation: enabled HTTP checks cover explicit request scope, while discovered ancestor traversal and local/direct paths have different boundaries. RT-RTE-memory-roundtrip, RT-BAP-retained-context, MEM-CLM-scope-visibility, RT-RTE-access-control.
Summary candidates only identify source windows. A caller must retrieve, author and separately append a summary or checkpoint. Stored Summarizes targets suppress already-covered windows without verifying summary adequacy or current validity of the summarizer. The report verifier also found a Compression-role convention using positional refs that become References, not that coverage relation. No complete comparison of compression bookkeeping follows. Limited reload and unchecked summaries afford continuity but do not support the advertised loss of no decisions. MEM-RTE-summary-selection, MEM-OBJ-summary-candidate, EPI-CLM-context-continuity, report verification limits.
Handoff delivery stops before task success
The representative cookbook workflow records an unfinished flaky-auth investigation, retrieves a Handoff-role/concept-filtered hit and appends the next agent's accepting checkpoint with a backlink. It preserves done, pending and next action. However, one function uses the same open chain and already holds the handoff index; it does not demonstrate restart or an independently starting agent resolving the target from retrieval. Assertions check a hit and reference, not the proposed fixed-clock rerun or a repaired application. Invalid refs can block admission; recovery remains caller-owned, and post-write failures require care before retry. The core skill's Summary/Checkpoint convention is not automatically interchangeable with this Handoff-role selector. WF-RTE-session-continuation.
For a scenario requiring retrievable state and explicit continuation markers, the implementation supplies meaningful coordination mechanisms. For one requiring lossless context recovery, verified diagnosis, automatic delegation or accepted application repair, the same evidence stops short. A persisted intention is delivered memory, not delivered task execution.
Instruction evolution is not yet demonstrated improvement
Skill uploads retain immutable versions through full bodies and patches; requested reads reconstruct instructions, with latest meaning newest rather than evaluated best. Service signatures are conditional on active keys and do not cover direct registry/dashboard admission. Revocation returns a separate status alongside readable content; authored warnings do not necessarily carry the entry's retirement reason. Host nonexecution and behavioral rollback depend on the consumer honoring status. RT-RTE-skill-admission, MEM-OBJ-skill-body, EPI-RTE-skill-withdrawal.
Together with numerical retrieval updates, this supports retained behavior-shaping state and available instruction evolution. The unresolved learning question is whether a retained criticism or update improves capacity for future action, not whether files can change. No such comparison was supplied. Theory-builder conditions remain separately unestablished for live uploaded theories: concrete localized theory content, decisions depending on that content, content-directed criticism and a next round shaped by its result are not demonstrated by version storage. Criticism of a consumed theory improving future capacity is likewise uninspected. Reflection's bidirectional self-representation and autonomy's role-by-role computational ownership are not established; automatic derivatives do not settle either property. These are independent uncertainties, not a combined negative verdict. EPI-CLM-fleet-improvement, RT-RTE-skill-admission, MEM-RTE-derived-update.
At the declared memory-system boundary, exercised self-improvement is unestablished over the claimed project horizon. There is no independently evaluated objective/update episode showing evidence caused update determination, the update changed MentisDB's own organization, that organization entered live later use, and subsequent operation causally depended on it. Host skill consumption and effects are outside captured execution evidence; release installation is operator-directed maintenance, not a memory-derived improvement. This does not deny the available evolution pathways or turn missing episodes into architectural absence. EPI-CLM-fleet-improvement, RT-RTE-release-update.
Shipped evaluators compare retrieved text with dataset evidence using bidirectional substring matching. Partial or empty strings can satisfy that predicate; LoCoMo broadens evidence with neighboring turns, and ConvoMem abstention measures missing evidence matches rather than an answer model refusing to answer. Diagnostics do not establish generated-answer accuracy, causal benefit or an automatic diagnostic-to-successor update. MEM-RTE-retrieval-evaluation.
The assessment would change with linked restart/compaction traces testing retained decisions, concrete extraction and reviewed-admission pairs, live skill criticism/update/consumption episodes against an independently specified objective, and controlled subsequent-task comparisons. Enforcing summary adequacy, preserving graph exclusions across transfers or checking revoked status at actual execution would change particular guarantees only on the routes covered. Source-read log: none; synthesis used the supplied members rather than adding source evidence.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Shipped compression conventions are incompletely catalogued; Compression-role refs become References rather than candidate-suppression Summarizes edges. | SRC-1; MEM-RTE-summary-selection, MEM-OBJ-thought-content, RT-BAP-core-skill, report verification limits | Read-only selector, caller checkpoint follow-on and verifier's cookbook compression example | Complete compression-form and coverage-bookkeeping comparisons; automatic or observed semantic compaction | Accepted records covering each convention's production, links, coverage and later reader; execution evidence for exercised compaction |
| The substrate profile groups vectors, indexes and overlays in one partial unit; local positive witnesses do not apply to every part. | SRC-1; MEM-OBJ-embedding-vector, MEM-OBJ-retrieval-state, MEM-RTE-derived-update, profile verification limits | Dense sidecars, rebuilt indexes and persisted overlays, with incomplete HNSW/WAL detail | Per-part substrate counts, exact co-location and complete backend inventories | Separate operative-part substrate records and classifications, including backend persistence detail |
| No observed host, provider, benchmark or update episode; external internals and deployed state are not captured. | SRC-1; boundary, RT-BAP-retained-context, EPI-CLM-context-continuity, EPI-CLM-fleet-improvement, WF-RTE-session-continuation | Source wiring and doctrine at the frozen commit; one-function handoff example | Actual compliance, independent restart, lossless continuation, repaired task, improved future capacity or exercised self-improvement | Registered traces linking retained inputs, consumer selection, actions and outcomes; independently specified objectives and causal contrasts for improvement |
| Live semantic content, caller approval and upstream provenance remain unresolved; relevant routes are not exhaustively assessed. | SRC-1; MEM-OBJ-thought-content, MEM-OBJ-skill-body, EPI-OBJ-extracted-thoughts, EPI-RTE-extraction-schema, MEM-RTE-summary-selection | Inspected append, extraction, summary, import, bundle and skill routes; no live artifacts | Complete semantic-curation or agency inventory; accepted ampliative knowledge, system-wide absence of acceptance, or theory-builder properties | Input/output and provenance-linked artifacts; evidence-consuming acceptance records; concrete theory, criticism and next-round consumer traces; inspection of relevant alternatives |
| Partial implementation inspection and differing enforcement envelopes remain. | SRC-1; RT-RTE-access-control, RT-RTE-memory-roundtrip, RT-RTE-state-maintenance, MEM-RTE-derived-update, MEM-RTE-chain-withdrawal | Shared append/open and selected transport/admin routes; incomplete cross-process, restore, exact/history, UI, integration and backend detail | Exhaustive security, isolation, crash durability, freshness, transaction rollback or secure-erasure assurance | Route-complete enforcement and storage contracts, backend/concurrency inspection and scoped failure/recovery evidence |
| Later diagnostic/log/callback consumers and mixed derivative inputs are incompletely known. | SRC-1; MEM-OBJ-retrieval-state, MEM-RTE-retrieval-evaluation, MEM-RTE-webhook-delivery, MEM-RTE-derived-update | Requested historical returns, current append-event dispatch and numerical derivative witness | Complete pull-only inventory, push-signal classification or complete trace-learning/source inventory; automatic diagnostic-driven semantic improvement | Accepted records tracing original inputs and every qualifying later update/consumer, including external callback retention where included |
| Evaluator predicates have limited licenses and no supplied run or intervention. | SRC-1; MEM-RTE-retrieval-evaluation | Three shipped retrieval matchers and diagnostic consumers | Measured answer quality, reference truth, benchmark superiority, causal component effects or transferable task benefit | Audited references, measured executions, task-answer evaluations and independently varied component comparisons |