Class CognitiveMemoryManager

This manager as one session sees it: the same store, graph and mechanisms, with a working memory of its own. encode, retrieve and assembleForPrompt on the view focus traces in that working memory, and the view's active context lists only the traces the session encoded or recalled. Sessions that call the manager directly share its one working memory, whose slot list (trace ids and activations) is part of every assembled context.

shutdown() on the view does nothing: whoever owns the manager shuts it down.

Stage E: typed-network store, or null when typed-network not configured.

Stage E: typed-network observer (LLM extractor), or null.

Stage E: typed spreading activation, or null (only set for 'full' variant).

Encode a new input into a memory trace. Called after each user message.

Retrieve relevant memories for a query. Called before prompt construction.

Assemble memory context for prompt injection within a token budget.

Feed a conversation message to the observation pipeline.

Pipeline flow:

1. Observer extracts typed observation notes from buffered messages
2. Notes are fed to the Reflector for consolidation into long-term traces
3. Reflected traces are encoded via encode() (typed as semantic/episodic/etc.)
4. Superseded traces are soft-deleted
5. Commitment and intention notes are auto-registered with ProspectiveMemoryManager

Check prospective memory triggers (Batch 2).

Register a new prospective reminder/intention.

List active prospective reminders.

Remove a prospective reminder.

Rehydrate a gisted/archived trace to its original verbatim content.

Delegates to the configured IMemoryArchive. Returns null when no archive is configured or when the trace is not found/integrity fails.

Run consolidation cycle (Batch 2).

Get memory health diagnostics.

Track a conversation message for context window management. Call for every user/assistant/system/tool message in the conversation.

Run context window compaction if needed. Call BEFORE assembling the LLM prompt. Returns the (potentially compacted) message list for the conversation. If infinite context is disabled, returns null (caller should use original messages).

Get the rolling summary chain text for prompt injection.

Get context window transparency stats.

Get full transparency report (for agent self-inspection or UI).

Get compaction history for audit/UI.

Search compaction history for a keyword.

Get the context window manager (for advanced usage).

Shutdown and release resources.

Access the underlying long-term memory store for diagnostics/devtools.

Delete every trace of one scope for good (MemoryStore.deleteScope), a batch at a time. What this manager keeps of each trace in a batch goes before the batch's traces (forgetTraces), so a call that fails part way can be made again and finds the same traces. With turnUserId, only the scope's traces that user filed as a second user on the session.

Throws

TypeError when scopeId is empty, or when turnUserId is given and is not a non-empty string.

Delete the given traces for good, as deleteScope does for a scope. An id the store finds no trace of (no Brain row, nothing it holds) is skipped, and nothing this manager keeps under that id is touched.

Re-files the Brain's plain user and persona scope ids that hold a /, a % or a : in the encoded form a learning layer reads, once per Brain, and moves what the manager's store holds of the traces it moved; on a Brain another handle or process re-filed, the store follows the Brain's rows (MemoryStore.ensureEncodedScopeIds). The layer calls it when it first opens the manager. Optional, so a host's own implementation keeps compiling.

Total number of memory traces currently resident in the manager's in-memory trace cache. Ergonomic passthrough to MemoryStore.getTraceCount; used by agentos-bench for memory-footprint telemetry without reaching into getStore().

Access the working-memory model for diagnostics/devtools.

Get the resolved cognitive-memory runtime config.

Get graph module when enabled.

Get observer module when enabled.

Step-8: accessor mirror of getObserver, for the reflector.

Step-8: Force the reflector to run over pending notes regardless of threshold. Encodes reflection traces, soft-deletes superseded IDs. Safe to call when no reflector or no pending notes exist (returns an empty result). Errors do not propagate — reflection is non-critical.

scopeOverride forces every encoded reflection trace to use the caller-supplied scope + scopeId, overriding whatever the reflector LLM invented. Callers that want all reflection traces to land in a single canonical scope (e.g. bench adapters that retrieve under user/bench) should pass this override.

Get prospective-memory manager when enabled.

Export the full brain state as a JSON string. Delegates to JsonExporter through the MemoryStore's brain. Throws if no brain is attached.

Import a JSON brain payload into the attached brain. Delegates to JsonImporter through the MemoryStore's brain. Throws if no brain is attached.

Attach a HyDE retriever to enable hypothesis-driven memory recall.

When set, the retrieve() and assembleForPrompt() methods can accept options.hyde = true to generate a hypothetical memory trace before searching. This improves recall for vague or abstract queries by producing embeddings that are semantically closer to stored traces.

Example

memoryManager.setHydeRetriever(new HydeRetriever({
  llmCaller: myLlmCaller,
  embeddingManager: myEmbeddingManager,
  config: { enabled: true },
}));
Copy

Get the HyDE retriever if configured, or null.

Return the attached neural reranker, or null when none is configured. Public read-only accessor for Step-3 bench wiring: the bench constructs a per-case HybridRetriever that needs the same reranker the manager uses, without bracket-accessing the private field.