Basic Memory synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

Basic Memory is a persistent knowledge engine with shipped assistant integrations, not the enclosing assistant's general-purpose runtime. It accepts Markdown notes, maintains derived access structures, and supplies selected retained content through tools and host callbacks. This account covers the canonical repository at a72751c5d626380dedeed30f55f7a80ba3dc710d, with an applicability cutoff of 2026-10-16. Its evidence is inspected implementation and doctrine/design from SRC-1, not target execution, production sessions or causal experiments. The boundary is complete artifact, partial loop: the artifact includes first-party callers and instructions, but their material routes are not exhaustively traced. See boundary and evidence and the Source register.

From a caller's proposal to accepted and published state

An enclosing user or assistant chooses a note operation. MCP write_note, also reachable through a CLI wrapper, resolves the project and passes a typed request to the knowledge API. Deterministic service code identifies the exact path owner and accepts or refuses the mutation in a database transaction. The caller still owns the next reasoning step. Server instructions ask assistants to obtain consent before creating notes; this is model-facing guidance, not a service requirement for per-content human approval. Enabled host capture can write automatically (RT-RTE-note-write, RT-BAP-server-instructions).

Replacement policy is conditional. An occupied path can be refused when overwrite is false; overwrite can be unconditional or checksum-conditioned. A collaboration-relay branch can supersede a mismatching revision under specified previous-source/storage-head conditions. Direct CRUD and filesystem paths do not inherit every path-addressed tool guard. Thus checksums provide scoped revision control, not a universal stale-write rejection guarantee (RT-RTE-note-write).

Acceptance, file publication and retrieval readiness are different completion points. Production local materialization queues an accepted snapshot, checks versions and existing bytes, publishes status, indexes successful writes and schedules vector work separately. Per-note queue serialization is process-local. Graceful shutdown drains work; startup recovery can re-drive pending, writing or failed rows. Neither this protocol nor its separate file check and write proves crash durability, cross-process serialization or atomic compare-and-swap against arbitrary filesystem writers (RT-RTE-note-materialization).

The retained body, synchronization metadata and published graph serve different consumers. Structured note reload can return accepted entity content before every derived representation agrees; its resource fallback tests for missing content, not an empty body. Parsed graph generations concern publication consistency, not factual warrant. This distinction matters when a caller needs the accepted text rather than a search-ready projection (MEM-OBJ-markdown-content, MEM-OBJ-sync-state, MEM-OBJ-graph-publication, MEM-RTE-note-reload).

From retained material to a later working context

Requested graph context selects indexed knowledge by identifier or search, traverses relations and applies page/related-result bounds. Compact mode omits observations and source prose; explicit note reads can retrieve bodies, metadata or line slices. Omitted material remains reachable through further calls. These are controls over particular returns, not a total assistant context budget or host compaction mechanism (RT-RTE-context-build, MEM-RTE-note-reload).

The representative multi-step workflow is Pi working-thread continuation. With direct capture selected, valid configuration, project mapping and workspace trust enable automatic callbacks. After the assistant settles, fixed code extracts current-branch user/assistant text, requires user content and sufficient text, formats an extractive checkpoint, then invokes the CLI with Markdown on stdin. Tool-role results and hidden/nontext blocks are not a complete evidentiary transcript in this branch. Default session/branch-derived titles permit checkpoint replacement; explicit titles change that overwrite behavior. The host still performs the user's task (RT-RTE-pi-checkpoint, EPI-OBJ-capture-draft, workflow account).

On a later eligible agent start, prompt/timeframe/project-filtered search selects up to five checkpoints, each with an excerpt clipped to 500 characters, for a displayed host message. The wrapper tells the assistant to treat recall as reference data, not instructions; model interpretation is not mechanically separated by that warning. Automatic operations and model-tool capture have trust predicates that the human capture command does not share. Those predicates are not an all-write approval gate or a sandbox (RT-RTE-pi-checkpoint, RT-BAP-recall-data).

Delivery remains weaker than verified continuation. The subprocess helper checks exit status and JSON syntax, not a successful-write discriminant; a capture notification therefore does not independently certify admission or completed publication. Capture failure warns without immediate retry. Recall consumes its automatic attempt before searching, so failure need not retry until a new session or explicit command. The supported endpoint is a context handoff, not demonstrated use or task success (workflow delivery account).

Other hosts materially differ. Python hook capture separates raw lifecycle metadata from authored checkpoints and emits bounded session briefs. Its archive sweep parses, moves and age-prunes metadata; it does not promote that metadata into knowledge. Tau instead asks an external host model to fold public session text and a prior handoff into a new durable handoff. Pending/confirmed receipts and identity/digest checks support uncertain-write reconciliation rather than blind retry. Tau requests rationale preservation and later reads prior handoffs, but fidelity remains unobserved. Hermes uses extractive first/last-turn summaries and query prefetch. OpenClaw caches task/recent-activity recall and returns it through systemPromptAddition; that placement does not turn reference content into standing instructions, and child-note creation does not establish child read-back (RT-RTE-hook-checkpoint, MEM-RTE-trace-archive, MEM-RTE-tau-handoff, MEM-RTE-hermes-continuity, MEM-RTE-openclaw-context).

These routes establish useful architectural capacity for continuing selected work across context boundaries. They do not substantiate literal universal recall: capture omissions, clipping, query selection, failure handling and host dependence remain consequential. Deletion also removes accepted/search/vector state without revoking already copied host context. Archive-first lifecycle guidance is not a universal deletion prohibition (EPI-CLM-never-forgets, MEM-RTE-note-withdrawal).

What checking and retained experience establish

Automatic Pi, hook, Tau and Hermes writes connect recorded conversational experience to durable context with later consumers. This supports the memory comparison's local wired trace-learning findings: here that term means an automatic trace-fed durable-context route, not improved capacity for future action. Tau additionally consumes its previous handoff during the next durable update. The unresolved question is whether these mechanisms improve later action, preserve the needed qualifications, or merely supply more context. No comparison in this evidence measures that benefit (RT-RTE-pi-checkpoint, RT-RTE-hook-checkpoint, MEM-RTE-tau-handoff, MEM-RTE-hermes-continuity, profile verification).

Basic Memory also implements narrow, genuine checking. Schema inference derives counts and proportions over supplied note representations and proposes field rules; frequency does not warrant the underlying assertions or compel schema adoption. Schema validation reports mode-relative structural conformity: warning-mode pass can coexist with a missing required field. Drift reports compare declared fields with observed usage. These responses are not factual acceptance certificates or inspected core write vetoes (EPI-RTE-schema-inference, EPI-RTE-schema-validation, EPI-RTE-schema-drift).

Research and reflection instructions prescribe more than storage. Research calls for current sources, uncertainty and user approval to save. Reflection calls for full-note review, durability selection, source links and preserving unresolved disagreements. They remain doctrine-only routes rather than observed evidence-consuming acceptance or installed autonomous recurrence. The benchmark separately generates answers from selected context and judges gold-fact coverage/noncontradiction or appropriate abstention. Verdicts affect metric accounting and can be rejudged; extra assertions are allowed unless contradictory. This is neither a comprehensive hallucination audit nor a production note-revision loop (EPI-RTE-research-note, EPI-RTE-reflect-notes, EPI-RTE-qa-answer, EPI-RTE-qa-judge).

The independent properties remain separate. No actual formulated theory route is established for assessing theory-builder conditions: localized theory content (condition 1), decision dependence on that content (2), content-directed criticism with blame and changed reliance (3), and retained criticism shaping another round (4) are each unestablished here, not evidenced absent system-wide. Nor is criticism of a consumed theory shown to improve future-action capacity. Reflection instructions and their operation records do not establish a bidirectional reflective system. Automatic callbacks establish local computational control, not autonomy over every internal theory-building operation. For self-improvement, retained context can reach later behavior, but the record establishes neither objective-responsive determination of an improvement-directed organizational update nor subsequent operation causally depending on that change. Exercised self-improvement and successful improvement are therefore unestablished at this boundary; this does not negate the wired continuation mechanisms (EPI-RTE-reflect-notes, RT-RTE-pi-checkpoint, MEM-RTE-tau-handoff).

Embedding adapters request inference, not parameter training, and configurable model names are not immutable weight identities. Their opaque provider behavior is a separate uncertainty from learning through retained prose (RT-CMP-fastembed-provider, RT-CMP-openai-provider, RT-CMP-litellm-provider).

For a checkpoint-continuation scenario, the supported result is a connected capture–retain–select–deliver architecture. For structurally checking notes, it is a mode-relative diagnostic capacity. For factual knowledge acceptance or improved future action, the evidence stops earlier. A traced successful-write check plus read-after-write confirmation would strengthen the Pi delivery contract; captured later use would establish activation. An objective-defined comparison with and without the continuity mechanism could assess benefit at that treatment level. Component attribution would require separate variation of the component. An evidence-consuming factual acceptance trace or an objective-responsive update followed by causally dependent operation would change the epistemic or self-improvement assessment, respectively. No source checkout reads or new evidence were added for this synthesis.

Limitations

limitation	affected SRC IDs and record citations	inspected boundary	conclusion prevented	evidence that would resolve it
Continuity inventory omits individually traced alternatives identified by report verification. Codex can request an agent-authored post-compaction handoff; the continuation prompt supplies another requested search/recent-activity path with 300-character topic excerpts and instructions to read more.	SRC-1; report-verification limit, RT-RTE-hook-checkpoint, RT-RTE-context-build, memory inventory	Verifier inspection of src/basic_memory/cli/commands/hook.py and src/basic_memory/mcp/prompts/continue_conversation.py; partial route inventory.	Exhaustive route counts, complete cross-host transformation comparison, or inference that every checkpoint uses the common extractive formatter. These alternatives do not establish successful checkpointing or another verified knowledge-production loop.	Accepted route records tracing those producers, admission, selection, delivery and later consumers; execution evidence for operation claims.
Source-only evidence supplies no deployed activation, reliability, semantic fidelity or benefit comparison.	SRC-1; boundary, RT-RTE-pi-checkpoint, MEM-RTE-tau-handoff, EPI-RTE-qa-judge	Frozen implementation/doctrine; external hosts, production corpora and live services uncaptured.	Successful continuation, measured accuracy, causal benefit or exercised self-improvement.	Captured runs linking retained artifacts to later actions, fidelity checks and objective-defined controlled comparisons.
Admission, storage, maintenance and consumer inventory remains partial, including import/watch/CI/cloud paths, caches/vector persistence, other tools, providers and host alternatives.	SRC-1; memory inventory, RT-RTE-note-write, MEM-RTE-openclaw-context, profile verification	Selected local write/read paths and distinct host witnesses; all profile axes retain partial coverage.	Whole-system guard, storage, authority, trace-origin or curation enumeration; cross-host equivalence, isolation, exhaustive negatives or fixed model identity.	Additional accepted records for unresolved paths, provider identities, backing stores and actual later consumers.
Acceptance and recovery do not prove completed durable publication or universal withdrawal.	SRC-1; RT-RTE-note-materialization, MEM-RTE-note-reload, MEM-RTE-note-withdrawal	Local asynchronous publication, selected reload and delete paths; storage/deployment contracts incomplete.	Crash correctness, synchronized file/index/vector freshness, atomic filesystem replacement or deletion of every copy.	Failure/recovery traces across storage boundaries, concurrency evidence and complete cleanup/cache/retraction paths.
Evaluation outputs and prescriptions do not establish factual acceptance or production improvement feedback.	SRC-1; EPI-RTE-schema-validation, EPI-RTE-research-note, EPI-RTE-reflect-notes, EPI-RTE-qa-judge, memory evaluator discussion	Structural reports, doctrine and benchmark scoring; observer effects and specific QA artifact-file linkage remain untraced.	Truth certification, exercised accepted conjectures, autonomous reflection, QA-to-file persistence inference or evaluator-to-production-update attribution.	Concrete evidence-consuming acceptance decisions, verified evaluator provenance/calibration, recorded persistence paths and consequential update consumers.
Pi notifications and reference wrappers provide weaker assurance than verified writes and safe model use.	SRC-1; workflow delivery account, RT-BAP-recall-data	Exit/JSON checks, callback ordering and clipped host-message output; host interpretation unobserved.	Treating a notification as verified publication, guaranteed retry, full evidentiary replay or enforced instruction/data separation.	Success-discriminant and read-back verification, retry traces, and captured consumer behavior or enforceable host separation controls.