Cognee synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Cognee is a memory platform that acquires content, processes it into graph and vector representations, and supplies retained material to later answer and maintenance computations. Its public claim includes retaining project context and distilling session lessons for reuse across runs. The inspected implementation supports that architecture, including a model-mediated lesson acceptance route; it does not demonstrate factual reliability or improved future performance. See EPI-CLM-durable-lessons.
This account uses SRC-1, the Cognee repository at commit f154b32a9e544dac73e7ddf2010327cf9988edcd, with an applicability cutoff of 2026-10-14. The boundary is complete artifact, partial loop: the repository is pinned, but not every shipped producer, maintainer and consumer is traced. Evidence consists of inspected implementation and doctrine/design, not target executions or causal comparisons. External model and database internals, separately distributed plugins, Cloud and arbitrary enclosing applications are excluded implementations. Shipped callers remain included; the Company Brain driver provides a separate saved-memory recall invocation, but no supplied execution proves its success. Evidence identities and layers are in the Source register. This synthesis adds no source reads.
Acquisition and retention
The ordinary local permanent-memory request supplies content to remember, which resolves user and writable dataset scope, calls add, then runs cognify. Code fixes the task progression; models perform selected extraction and summarization, rather than choosing the pipeline's next task. Generated nodes, edges and embeddings cross graph/vector storage interfaces. Blocking errors can propagate; background processing reports through a running result. Remote configuration delegates under a separate server contract. See RT-RTE-permanent-memory.
Retention is not a single state transition. Session remember first writes cache data; optional automatic improvement is a separately scheduled continuation. Its failure does not undo the session write. Typed QA, trace and feedback branches return after their own writes, while typed SkillRuns are graph-backed. A successful session-storage result therefore does not establish permanent graph publication. Automatic maintenance depends on configuration, host admission and debounce; the host callback fails open and explicit improve bypasses it. Process-local claims limit overlapping maintenance, not cross-worker contention or eventual processing. See RT-RTE-session-improve, MEM-RTE-typed-admission, RT-CLM-auto-admission and RT-CLM-improve-coordination.
The retained products have distinct roles. Graph prose supplies content; graph identifiers and provenance govern traversal and cleanup; numerical indexes support retrieval. Session records retain question/answer text, sanitized host traces, feedback and derived guidance. Skills separate deployed procedure text, discovery/scope metadata, run summaries and proposed revisions. Sharing storage does not make these interchangeable. Neither dense retained representations nor scalar feedback weights establish provider-weight training. See MEM-OBJ-graph-content, MEM-OBJ-graph-structure, MEM-OBJ-vector-index, MEM-OBJ-session-agent-trace and MEM-OBJ-skill-run.
Maintenance, checking and admission
Session turns can generate guidance candidates and ratings. Deterministic gates reject invalid, oversized, low-confidence or duplicate guidance. Later ranking uses content, confidence, helpfulness and recency; default guidance budgets fit whole content into 400 characters per section and 1,200 characters overall, not a total conversation-token bound. Trace promotion separately converts QA or trace text into graph content and extracts active agent-profile lessons. Oldest-pending graph persistence and newest-at-most-40 lesson extraction have different coverage; an extraction watermark can advance past older unexamined backlog. Agent-profile storage also does not imply injection into the inspected agentic guidance block, which defaults to QA profile. See MEM-RTE-session-guidance and MEM-RTE-trace-promotion.
Distillation adds a substantive acceptance architecture. Gated entries and bounded QA timelines reach a curator, which proposes durable lessons. A writer judges support, novelty and durability against resolved member entries, retrieved prior lessons and a glossary. Code grants publication eligibility only to nonfailed, accepted, nonblank statements. Accepted text and optional why_learned prose become documents, then enter add and cognify for graph organization and later reuse. Checking, admission, retention and integration are separate functions. See MEM-RTE-session-distillation, EPI-RTE-lesson-check, EPI-RTE-lesson-admission, EPI-RTE-lesson-retention and EPI-RTE-lesson-integration.
That acceptance licenses a model judgment for graph reuse, not independent corroboration. The writer lacks the original QA transcript, unresolved member IDs are omitted, and novelty/glossary lookup failures produce empty results. No actual input/output and decision instance establishes whether a generated lesson preserves, entails or extends its evidence. Imported source warrant remains unknown. Publication can survive a later failed model call; failed calls leave inputs retryable, whereas successful rejection or empty proposals can advance the watermark. See EPI-RTE-lesson-curation, EPI-RTE-lesson-check and EPI-RTE-distillation-recovery.
Other maintenance changes reliance without establishing truth. Feedback and preference updates alter ranking state or durable preference text. Optional “truth subspace” geometry uses accepted-learning embeddings to rerank chunks; matching integer epochs suppress some stale coordinates but cannot exclude a documented abandoned-basis/reused-epoch countercase. Optional global summaries compress child content for a later query prelude. Contradiction checks retain model warning edges without vetoing ingestion; temporal supersession records current-by-recency policy, not empirical verification of the latest assertion. Their later reader policies remain unresolved. See MEM-RTE-ranking-maintenance, MEM-RTE-truth-maintenance, MEM-RTE-summary-readback, EPI-RTE-contradiction-annotation and EPI-RTE-temporal-supersession.
Later consumption and a representative workflow
Recall returns selected session, graph or other lane results. Caller-pinned search types override automatic routing; session success can short-circuit graph work in auto-fallthrough mode. Without a usable LLM, the default graph path chooses chunks rather than generated completion. The decorator offers another route: it selects recent trace feedback and graph context within a 4,000-character envelope, invokes the wrapped function, then attempts trace persistence. Only structured-output gateway calls within its active frame automatically receive that memory. The host still owns its actions, and the decorator's graph selection must not be assumed embedding-based merely because an index exists. See RT-RTE-recall-dispatch, RT-RTE-decorated-agent, its memory annotation, and profile verification.
The representative multi-step workflow is explicit agentic completion over one dataset, not default recall. Code prepares graph context, optional history/guidance, a skill catalog and a filtered tool manifest. The model chooses successive AgentStep actions or a final answer. Requested load_skill selects from the active map and supplies procedure text; a 12,000-character body cap is followed by an 8,000-character tool-result cap, so full stored procedure delivery is not guaranteed. Tool-scope and dataset-permission checks precede in-process handler invocation; they do not sandbox its effects. See RT-RTE-agentic-tools, RT-OBJ-agent-step and WF-RTE-skill-loading.
Caught scope, permission and invocation failures become error text for the model's next decision. Other exceptions can escape; previous tool effects are not rolled back. A nonblank final answer wins immediately. Iterative exhaustion or no actionable step causes an additional best-effort completion, beyond the turn bound. SkillRun and QA recording are best effort and selective, not complete replay. The inherited evidence-append call is a no-op, not final-answer validation. Delivery means returning the answer list, not verified task success or host adoption. This is a supported possible progression, not an observed run; a request may use no tools. See the workflow account and MEM-OBJ-skill-run.
Adaptation and independent property assessments
The strongest supported adaptive contribution is implemented trace-fed production of later-consumed guidance, lessons and ranking state. Recorded turns and host feedback can shape durable updates without updating model weights. This establishes update wiring, not improved capacity for future action. Skill revision adds failure-selected drafting, retained old/new procedures and rationale, and caller-selected installation into future skill delivery. Installation checks identity, not performance improvement or stale-base consistency. The report verifier also found a shipped model-feedback demo that writes a scored SkillRun and automatically applies a proposal; the inventory therefore must not imply that all callers use neutral scores or that application is necessarily human-controlled. No demo execution or subsequent revised-skill use is supplied. See MEM-RTE-trace-promotion, MEM-RTE-session-distillation, MEM-RTE-ranking-maintenance, RT-RTE-skill-revision and report verification.
For the inspected proposal mechanism, RT-RTE-skill-revision supports separate conclusions, not a ladder:
| Property | Supported status and boundary |
|---|---|
| Theory-builder condition 1: localized formulation | Afforded at whole-procedure granularity; no supplied procedure instance establishes a formulated theory or separately revisable assumptions. |
| Condition 2: decisions depend on theory content | Afforded through progressive guidance delivery; actual content-dependent decisions are unobserved. |
| Condition 3: content-directed criticism | Uninspected; failure scores and requested revisions do not establish a stated challenged claim, result and placement of blame. The additional demo requests missing-instruction criticism but supplies no performed instance. |
| Condition 4: criticism shapes another round | Afforded through retained/applied guidance and future run evidence; actual criticism retention and consequential reuse are unobserved. |
| Learning from criticism of a consumed theory | Uninspected; no capacity improvement or attribution evidence is supplied. |
| Reflection | Afforded: stored guidance and run outcomes represent selected operation and can mediate revision; actual bidirectional operation is unobserved. |
| Reflective theory builder | Uninspected; all four conditions are not demonstrated for criticism of method text against its own operation. |
| Autonomous theory builder | Uninspected; the core apply caller's role is unresolved and the additional automatic demo does not establish every internal theory-building role. |
| Self-improvement | Afforded, within the skill-guided retriever across revision and a possible later invocation, with better skill task performance as the independent objective. Failure evidence can shape replacement and later instruction supply, but subsequent behavior causally depending on that change is unobserved. Successful improvement is not established. |
These unresolved properties do not negate the positive update mechanisms. Offline golden-answer evaluation is implemented separately, but no inspected linkage makes its scores a production lesson or skill-install acceptance gate. No controlled comparison identifies either an overall improvement or a component effect. See MEM-CLM-evaluation-boundary and EPI-RTE-benchmark-evaluation.
Scenario-relative assessment
For cross-invocation memory supply, Cognee has source-supported production and later-consumer routes, including shipped callers. For evidence-sensitive lesson reuse, it adds explicit model judgment with publication force, but incomplete evidence views and unobserved fidelity limit the warrant. For bounded tool work, it owns action dispatch and iteration policy, not arbitrary host execution or isolated effects. For comprehensive withdrawal, it follows recorded dependency chains rather than all semantic consequences: traces without graph IDs, append-only alternatives and unresolved derived-state retraction prevent universal erasure. See MEM-RTE-memory-withdrawal.
Assessment would change with linked production/update/consumer traces, actual lesson decision pairs, complete score-producer and apply-driver coverage, and controlled future-task comparisons. Exact basis identity checked per truth-coordinate would address the epoch countercase; complete derivative dependency and backend deletion tracing would strengthen withdrawal conclusions. These are named evidence or system changes, not inferred benefits or adoption recommendations.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Partial shipped-loop and interface coverage | SRC-1; RT-RTE-permanent-memory, RT-RTE-recall-dispatch, boundary | Central APIs and selected callers; frontend, connectors, workers, alternate lanes and deployment paths incompletely traced | Whole-system closure, uniform transport behavior, universal authorization or system-complete negatives | Material producer/maintainer/consumer traces across those included alternatives and deployed contracts |
| No observed operation or causal comparison | SRC-1; RT-RTE-agentic-tools, RT-RTE-skill-revision, EPI-RTE-benchmark-evaluation | Implementation and doctrine/design only | Successful persistence, activated guidance, reliability, beneficial learning or component-causal effect | Linked run evidence and objective-specific comparisons with appropriate independent variation |
| External inference identity and source provenance unresolved | SRC-1; RT-CMP-llm-gateway, RT-CMP-embedding-config, MEM-RTE-typed-admission | Dispatch/configuration and typed writes; provider internals and original mixed payloads unresolved | Immutable model attribution, authenticated trace origins or factual source warrant | Immutable provider/index provenance and origin-linked retained inputs |
| Lesson writer evidence and semantics incomplete | SRC-1; EPI-RTE-lesson-curation, EPI-RTE-lesson-check, EPI-RTE-lesson-admission | Bounded curator view; writer omits original QA and unresolved IDs; searches fail open | Reliable support/novelty judgments, observed accepted ampliative claims or warranted derivation | Actual input/output/decision pairs, complete member evidence and independently checked support |
| Maintenance completion is not all-window publication | SRC-1; RT-RTE-session-improve, MEM-RTE-trace-promotion, RT-CLM-improve-coordination | Fatal-stage branch, nonpropagated window errors, capped latest extraction and process-local claims | All-QA admission, all-trace extraction, cross-worker exclusion or eventual recovery | End-to-end stage aggregation, backlog/failure traces and shared coordination contract |
| Context delivery is route-specific and bounded only locally | SRC-1; MEM-RTE-session-guidance, RT-RTE-decorated-agent, WF-RTE-skill-loading, profile verification | Profile-specific guidance, structured-output injection and dual skill caps; decorator graph selector unresolved | Universal agent-lesson injection, embedding selection on the decorator path, complete procedure reception, total token bound or obedience | Actual selector trace, all consumer paths, delivered-context captures and content-dependent action evidence |
| Conflict metadata lacks established later consumer policy | SRC-1; EPI-RTE-contradiction-annotation, EPI-RTE-temporal-supersession | Annotation and recency-tag producers | Universal abstention, withdrawal, latest-only answering or factual endorsement | Traced downstream conflict/supersession consumers and evidence-sensitive decisions |
| Truth epoch is not exact basis identity | SRC-1; MEM-RTE-truth-maintenance, EPI-RTE-truth-reranking, EPI-CLM-epoch-recovery | Write-before-commit ordering and documented reused-epoch countercase | Unconditional coordinate consistency or factual-verification claim | Exact basis validation per chunk and failure-path evidence; factual verification needs a separate criterion |
| Withdrawal and reconstruction do not close derivative lifecycles | SRC-1; MEM-RTE-memory-withdrawal, MEM-RTE-ranking-maintenance, MEM-RTE-summary-readback | Recorded graph-to-session dependencies and selected deletion helpers; cache/backend alternatives partial | Universal erasure, tenant extent of all-cache prune, inverse ranking updates or complete summary/preference/lesson/procedure retraction | Adapter and derivative lifecycle traces, dependency coverage and tenant-scoped deletion evidence |
| Skill-feedback inventory omits a shipped model-feedback driver | SRC-1; RT-RTE-skill-revision, MEM-OBJ-skill-run, report verification | Core revision/recorder account; verifier inspected additional feedback demo and its instruction | Complete comparison of score provenance, computational admission and criticism-producing workflows; generalizing neutral scores to every caller | Canonical coverage of the demo and other producers; execution/decision instances and later revised-skill use for stronger property conclusions |
| Proposal installation and evaluator linkage lack performance acceptance evidence | SRC-1; RT-RTE-skill-revision, MEM-CLM-evaluation-boundary, EPI-RTE-benchmark-evaluation | Target-checked replacement and separate reference-answer reporting | Tested improvement, complete theory-building autonomy or exercised self-improvement | Role-by-role update chain, stated criticism, retained consequential reuse, linked acceptance tests and future-task comparison |
Report verification declares the inventory omission above and no blocker; profile verification declares no remaining limit or blocker. Reconciliation establishes no unresolved cross-report conflict. The remaining uncertainty is carried at its affected mechanism, not converted into absence or a system-wide grade.