Skip to content

api.md

Latest commit

 

History

History
704 lines (492 loc) · 29.9 KB

api.md

File metadata and controls

704 lines (492 loc) · 29.9 KB
Raw

API Reference

Standalone CLI Commands

The canonical CLI is remnic. The legacy engram binary remains as a forwarder during the rename window, and the standalone commands below are still available through either name.

Command Description
remnic init Create remnic.config.json in the current directory
remnic status [--json] Show server/daemon status and health
remnic query <text> [--json] [--explain] Query memories; --explain shows per-tier latency breakdown
remnic doctor Run diagnostics (Node version, config, API key, memory dir, daemon status)
remnic config Show current resolved configuration
remnic daemon <start|stop|restart> Manage the background HTTP server
remnic tree <generate|watch|validate> Workspace context tree operations
remnic onboard [dir] [--json] Analyze a project directory (language detection, doc discovery, ingestion plan)
remnic curate <path> [--json] Curate files into memory with duplicate/contradiction detection
remnic review <list|approve|dismiss|flag> [id] Review inbox management
remnic sync <run|watch> [--source <dir>] Diff-aware filesystem sync
remnic dedup [--json] Find duplicate memories
remnic connectors <list|install|remove|doctor> [id] Manage host adapter connectors
remnic space <list|switch|create|delete|push|pull|share|promote|audit> Manage personal, project, and team memory spaces
remnic benchmark <run|check|report> [queries...] [--explain] [--baseline=<path>] [--report=<path>] Run benchmarks, check for regressions, generate reports
remnic versions <list|show|diff|revert> <page-path> [version-id(s)] Page version history management
remnic taxonomy <show|resolver|add|remove> [args] MECE taxonomy knowledge directory management
remnic enrich <entity-name|--all|audit|providers> [--dry-run] External entity enrichment pipeline
remnic binary <scan|status|run|clean> [--dry-run|--force] Binary file lifecycle operations

All commands accept --json for machine-readable output. The CLI resolves configuration from:

  1. REMNIC_CONFIG_PATH environment variable (ENGRAM_CONFIG_PATH is still accepted in v1.x)
  2. ./remnic.config.json in the current directory
  3. ~/.config/remnic/config.json (default)

See the Platform Migration Guide for detailed setup and usage instructions.

Universal Access Layer

Remnic exposes one shared local service layer through both HTTP and MCP adapters. The HTTP server is bearer-token protected by default and binds to loopback unless you override agentAccessHttp.host.

HTTP

Core routes:

  • GET /engram/v1/health — service health plus projection/search availability
  • POST /engram/v1/recall — shared recall entrypoint
  • POST /engram/v1/recall/explain — last recall snapshot plus intent/graph debug state
  • POST /engram/v1/memories — explicit memory write path
  • POST /engram/v1/suggestions — queue review-first memory suggestions
  • GET /engram/v1/memories — browse memories with query/status/category filters
  • GET /engram/v1/memories/:id — fetch one memory
  • GET /engram/v1/memories/:id/timeline — fetch one memory lifecycle timeline
  • GET /engram/v1/entities — list entities
  • GET /engram/v1/entities/:name — fetch one entity
  • GET /engram/v1/review-queue — latest governance review bundle when present
  • GET /engram/v1/maintenance — health plus latest governance artifact summary
  • GET /engram/v1/trust-zones/status — trust-zone store status, counts, and latest record summary
  • GET /engram/v1/trust-zones/records — browse trust-zone records with zone/source/query filters
  • POST /engram/v1/trust-zones/promote — dry-run or apply a trust-zone promotion
  • POST /engram/v1/trust-zones/demo-seed — explicitly seed an opt-in buyer demo dataset
  • POST /engram/v1/review-disposition — operator review decision write path
  • POST /engram/v1/observe — feed conversation messages into LCM archive and extraction pipeline
  • POST /engram/v1/lcm/search — full-text search over LCM-archived conversations
  • POST /engram/v1/lcm/compaction/flush — drain pending LCM observations before a host compaction
  • POST /engram/v1/lcm/compaction/record — record a completed host compaction checkpoint
  • GET /engram/v1/lcm/status — LCM availability and stats
  • POST /v1/citations/observed — Record observed citation usage for attribution tracking

Recall request fields:

  • query (required)
  • sessionKey
  • namespace
  • topK
  • mode (auto, no_recall, minimal, full, graph_mode)
  • includeDebug
  • cwd (string, optional) — absolute path to the working directory. When provided and no coding context exists for the session, the server resolves git context automatically (see Coding agent mode).
  • projectTag (string, optional) — project name (e.g. "blend-supply"). Creates a tag:<name> coding context. Takes precedence over cwd when both are provided.

Recall response fields:

  • results
  • count
  • traceId
  • plannerMode
  • fallbackUsed
  • sourcesUsed
  • budgetsApplied
  • latencyMs

Write request envelope:

  • schemaVersion
  • idempotencyKey
  • dryRun

Write endpoints share the same explicit-capture validation and duplicate suppression as the OpenClaw tooling, enforce request-size limits, and are rate-limited before mutation paths run.

Trust-zone routes

GET /engram/v1/trust-zones/status

  • returns { namespace, status }
  • status.records.byZone shows quarantine/working/trusted counts
  • when poisoning defense is enabled, trust-score bands and aggregate provenance scores are included

GET /engram/v1/trust-zones/records

Query parameters:

  • q — free-text search over summary, tags, entity refs, and metadata
  • zone — quarantine, working, or trusted
  • kind — memory, artifact, state, trajectory, or external
  • sourceClass — tool_output, web_content, subagent_trace, system_memory, user_input, or manual
  • limit
  • offset
  • namespace

Each returned record includes:

  • provenance summary (sourceClass, sourceId, evidenceHashPresent, anchored)
  • trust score details when poisoning defense is enabled
  • next-step promotion readiness (nextPromotionTarget, nextPromotionAllowed, nextPromotionReasons)
  • corroboration counts for risky working -> trusted promotions

POST /engram/v1/trust-zones/promote

Request fields:

  • recordId (required)
  • targetZone (required; working or trusted)
  • promotionReason (required)
  • recordedAt
  • summary
  • dryRun
  • namespace

POST /engram/v1/trust-zones/demo-seed

Request fields:

  • scenario (optional, default: enterprise-buyer-v1; also supports agentic-commerce-v1)
  • recordedAt (optional base ISO timestamp for demo records)
  • dryRun
  • namespace

This route is intentionally explicit and never runs automatically. Use it only when you want seeded demo data in the selected namespace. agentic-commerce-v1 is the synthetic commerce walkthrough for buyer preferences, exclusions, shipping urgency, and ask-before-checkout boundaries.

POST /engram/v1/observe

Feed conversation messages into the memory pipeline (LCM archive + extraction).

Request fields:

  • sessionKey (string, required) — conversation session identifier
  • messages (array, required) — array of { role: "user" | "assistant", content: string } objects; must be non-empty
  • messages[].sourceFormat (string, optional) — source payload format; supports openai, anthropic, openclaw, pi, lossless-claw, and remnic
  • messages[].parts (array, optional) — structured tool/file/message parts used by coding-agent integrations
  • namespace (string, optional) — target namespace; defaults to the resolved namespace from the principal
  • skipExtraction (boolean, optional) — when true, messages are archived in LCM but not sent through extraction
  • cwd (string, optional) — absolute path to the working directory. When provided and no coding context exists for the session, the server resolves git context automatically (see Coding agent mode).
  • projectTag (string, optional) — project name (e.g. "blend-supply"). Creates a tag:<name> coding context. Takes precedence over cwd when both are provided.

Response (HTTP 202):

  • accepted — number of messages accepted
  • sessionKey — echo of the session key
  • namespace — resolved namespace
  • lcmArchived — whether messages were archived in LCM
  • extractionQueued — whether messages were queued for extraction

Rate-limited to 30 requests per minute. See the Standalone Server Guide for details.

POST /engram/v1/lcm/search

Full-text search over LCM-archived conversation messages.

Request fields:

  • query (string, required) — search query
  • sessionKey (string, optional) — filter results to a specific session
  • namespace (string, optional) — filter by namespace
  • limit (number, optional, default: 10) — maximum results

Response (HTTP 200):

  • query — echo of the search query
  • namespace — resolved namespace
  • results — array of { sessionId, content, turnIndex } objects
  • count — number of results returned
  • lcmEnabled — whether LCM is enabled; if false, results will be empty

POST /engram/v1/lcm/compaction/flush

Drain pending LCM observation work for a session before a host compacts its local context. Pi uses this from session_before_compact so Remnic has the latest turns before the compacted checkpoint is generated.

Request fields:

  • sessionKey (string, required) — conversation session identifier
  • namespace (string, optional) — target namespace

Response (HTTP 200):

  • enabled — whether LCM is enabled
  • flushed — whether a flush was performed
  • sessionKey — echo of the session key
  • namespace — resolved namespace
  • reason (optional) — present when LCM is disabled

POST /engram/v1/lcm/compaction/record

Record the token delta for a completed host compaction. This lets Remnic correlate host-side compaction events with LCM checkpoints and later search/recall behavior.

Request fields:

  • sessionKey (string, required) — conversation session identifier
  • namespace (string, optional) — target namespace
  • tokensBefore (integer, required) — non-negative token count before compaction
  • tokensAfter (integer, required) — non-negative token count after compaction

Response (HTTP 200):

  • enabled — whether LCM is enabled
  • recorded — whether the compaction event was recorded
  • sessionKey — echo of the session key
  • namespace — resolved namespace
  • reason (optional) — present when LCM is disabled

GET /engram/v1/lcm/status

Returns LCM availability and statistics.

Response (HTTP 200):

  • enabled — whether LCM is enabled
  • archiveAvailable — whether the LCM archive is accessible
  • stats (optional) — { totalTurns } when LCM is enabled

POST /v1/citations/observed

Record that cited memories were used by the agent. Used for citation attribution tracking.

Request fields:

  • sessionId (string, optional) — Session identifier
  • namespace (string, optional) — Target namespace
  • citations (object, required) — Citation data containing:
    • entries (array, optional) — Array of { path: string, lineStart: number, lineEnd: number, note?: string } objects
    • rolloutIds (string[], optional) — Rollout IDs from the oai-mem-citation block

Response (HTTP 200):

  • ok (boolean) — Whether the request succeeded
  • submitted (number) — Number of citation entries submitted
  • matched (number) — Number of entries matched to existing memories
  • entriesReceived (number) — Number of citation entries in the request
  • rolloutIdsReceived (number) — Number of rollout IDs in the request

X-Engram-Principal Header

When the server is started with --trust-principal-header, requests can include an X-Engram-Principal header to override the authenticated principal for that request. This determines namespace read/write access. Without --trust-principal-header, the header is silently ignored.

MCP

Run the server with:

openclaw engram access mcp-serve

Available MCP tools:

  • remnic.recall — accepts optional cwd and projectTag for automatic project detection
  • remnic.recall_explain
  • remnic.memory_get
  • remnic.memory_timeline
  • remnic.memory_store
  • remnic.suggestion_submit
  • remnic.entity_get
  • remnic.review_queue_list
  • remnic.observe — accepts optional cwd and projectTag for automatic project detection
  • remnic.lcm_search
  • remnic.lcm_compaction_flush
  • remnic.lcm_compaction_record
  • remnic.day_summary
  • remnic.set_coding_context — attach or clear a session's coding context; accepts a full codingContext object or a projectTag shorthand

The legacy engram.* aliases remain available through the v1.x compatibility window.

The MCP adapter calls the same EngramAccessService methods used by HTTP, so equivalent request classes return the same structured payloads.

remnic.observe

Feed conversation messages into Remnic's memory pipeline (LCM archive + extraction).

Parameters:

  • sessionKey (string, required) — conversation session identifier
  • messages (array, required) — array of { role: "user" | "assistant", content: string } objects
  • messages[].sourceFormat (string, optional) — source payload format, including pi
  • messages[].parts (array, optional) — structured tool/file/message parts
  • namespace (string, optional) — target namespace
  • skipExtraction (boolean, optional) — skip extraction, archive in LCM only
  • cwd (string, optional) — absolute working directory path for automatic git context resolution
  • projectTag (string, optional) — project name for non-git sessions (creates a tag:<name> coding context)

Returns: { accepted, sessionKey, namespace, lcmArchived, extractionQueued }

remnic.lcm_search

Search the LCM conversation archive for matching content using full-text search.

Parameters:

  • query (string, required) — search query
  • sessionKey (string, optional) — filter to a specific session
  • namespace (string, optional) — filter by namespace
  • limit (number, optional) — max results to return

Returns: { query, namespace, results: [{ sessionId, content, turnIndex }], count, lcmEnabled }

remnic.lcm_compaction_flush

Flush pending LCM observation work before a host-side context compaction.

Parameters:

  • sessionKey (string, required) — conversation session identifier
  • namespace (string, optional) — target namespace

Returns: { enabled, flushed, sessionKey, namespace, reason? }

remnic.lcm_compaction_record

Record a host-side compaction event after the host has produced the compacted checkpoint.

Parameters:

  • sessionKey (string, required) — conversation session identifier
  • namespace (string, optional) — target namespace
  • tokensBefore (integer, required) — non-negative token count before compaction
  • tokensAfter (integer, required) — non-negative token count after compaction

Returns: { enabled, recorded, sessionKey, namespace, reason? }

remnic.day_summary

Generate a structured end-of-day summary from memory content.

Parameters:

  • memories (string, optional) — Pre-collected memory text; when omitted or empty, auto-gathers today's facts and hourly summaries from storage
  • sessionKey (string, optional) — Session identifier
  • namespace (string, optional) — Target namespace

Returns: Structured summary of the day's memory activity.

MCP over HTTP

The HTTP server also exposes an MCP JSON-RPC endpoint at POST /mcp, allowing remote MCP clients (e.g., Codex CLI, Claude Code) to use Engram tools over HTTP instead of STDIO:

openclaw engram access http-serve --host 0.0.0.0 --port 4318 --token "$TOKEN"

Clients send standard MCP JSON-RPC requests to http://<host>:4318/mcp with an Authorization: Bearer <token> header. Advertised MCP tools include both canonical remnic.* names and legacy engram.* aliases where supported. Write operations (engram.memory_store, engram.suggestion_submit, engram.observe, engram.lcm_compaction_flush, engram.lcm_compaction_record) are rate-limited consistently with the REST write endpoints - dry runs and idempotency replays do not count toward the limit.

Namespace-enabled deployments: If you have namespacesEnabled: true, pass --principal <name> to set the authenticated principal for all MCP connections. The principal must appear in writePrincipals for the target namespace. Without --principal, the principal resolves to "default", which may not have write access:

openclaw engram access http-serve --host 0.0.0.0 --principal generalist --token "$TOKEN"

Deployments with namespacesEnabled: false (the default) do not need --principal — all writes are permitted.

Agent Tools

These tools are registered with the OpenClaw gateway and are callable by agents.

memory_search

Search memories by semantic similarity.

Parameters:

  • query (string, required) — The search query.
  • limit (number, optional, default: 10) — Max results to return.
  • category (string, optional) — Filter by memory category.
  • namespace (string, optional) — Filter by namespace.
  • collection (string, optional) — QMD collection override for direct MCP/access calls.

When namespaces are enabled, unqualified searches use the authenticated principal's readable recall namespaces. Passing collection: "global" remains ACL-scoped to those readable namespaces; it does not bypass namespace isolation. Namespace-derived collection names are accepted only when they match a readable requested namespace. Arbitrary custom collections are rejected in namespace mode because Remnic cannot prove they are namespace-safe. Deployments without namespaces may still search a named custom QMD collection directly.

Returns: Array of matching memories with scores, paths, and content snippets.

memory_store

Manually store a memory without going through the extraction pipeline.

Parameters:

  • content (string, required) — The memory content.
  • category (string, required) — One of: fact, preference, correction, entity, decision, relationship, principle, commitment, moment, skill.
  • confidence (number, optional, default: 0.9) — Confidence score 0–1.
  • tags (string[], optional) — Tags to attach.

memory_store shares the same explicit-capture validation, sanitization, duplicate handling, lifecycle logging, and review-queue fallback used by memory_capture.

Returns: The stored memory's ID and file path, or the duplicate/review item identifier when Engram suppresses a direct write.

memory_capture

Create a structured explicit memory note that obeys captureMode policy.

Prefer this tool over inline notes when tool use is available. In explicit mode it is the primary write path; in hybrid mode it bypasses buffering and persists immediately when validation passes.

Parameters:

  • content (string, required) — One durable fact, decision, correction, commitment, or other standalone note.
  • category (string, optional, default: fact) — One of: fact, preference, correction, entity, decision, relationship, principle, commitment, moment, skill, rule.
  • confidence (number, optional, default: 0.95) — Confidence score 0–1.
  • namespace (string, optional) — Requested namespace, subject to namespace policy.
  • tags (string[], optional) — Tags to attach.
  • entityRef (string, optional) — Related entity id.
  • ttl (string, optional) — ISO timestamp or relative duration like 30m, 12h, 7d, or 2w.
  • sourceReason (string, optional) — Human/operator rationale recorded in lifecycle metadata.

Validation rules:

  • content must be 10–4000 chars
  • nested <memory_note> blocks are rejected
  • unsafe categories, secrets, credentials, and invalid namespace targets are rejected
  • exact duplicates are suppressed before write

If a direct write is rejected, Engram queues a sanitized pending_review memory instead of silently dropping the request.

Returns: The accepted memory id, duplicate target id, or queued review item id.

memory_profile

Retrieve the current behavioral profile.

Parameters: None.

Returns: The contents of profile.md.

memory_entities

List all tracked entities.

Parameters:

  • type (string, optional) — Filter by entity type (person, company, project, place).

Returns: Array of entity summaries with names, types, and fact counts.

memory_promote

Promote a memory to a shared namespace so other agents can access it.

Parameters:

  • memoryId (string, required) — The ID of the memory to promote.
  • targetNamespace (string, optional, default: shared) — Destination namespace.

Returns: The new path in the shared namespace.

memory_feedback

Record explicit feedback on a recalled memory.

Parameters:

  • memoryId (string, required) — The ID of the memory.
  • signal (string, required) — One of: thumbs_up, thumbs_down.
  • note (string, optional) — Optional explanation.

Returns: Confirmation with updated memory status.

memory_action_apply

Record a memory-action telemetry event with optional safe dry-run mode.

Parameters:

  • action (string, required) — One of: store_episode, store_note, update_note, create_artifact, summarize_node, discard, link_graph.
  • outcome (string, optional, default: applied) — One of: applied, skipped, failed.
  • reason (string, optional) — Operator rationale or note.
  • memoryId (string, optional) — Targeted memory ID if applicable.
  • namespace (string, optional) — Namespace to write telemetry into.
  • sourcePrompt (string, optional) — Prompt text used only for hash telemetry.
  • dryRun (boolean, optional, default: false) — Validate/report action without persisting telemetry.

Returns: Confirmation text; in dry-run, reports what would be recorded.

action_confidence

Return a read-only interruption-budgeting decision: ask, draft, act, refuse, or escalate.

HTTP: POST /remnic/v1/action-confidence or POST /engram/v1/action-confidence

MCP: remnic.action_confidence or engram.action_confidence

Parameters:

  • confidence (number, optional) - Overall confidence score 0-1.
  • risk (string, optional) - One of: low, medium, high, irreversible, restricted.
  • contextReadiness (string, optional) - One of: none, partial, sufficient.
  • retrievedMemories (array, optional) - Provenance/safety summaries for recalled memories.
  • currentContextScopes (array, optional) - Current user-context scopes.
  • userRules (array, optional) - Matched ask-before, do-not-use-outside-this-context, never, or requires-escalation rules.

Returns: The decision, confidence, blockers, reasons, factor breakdown, and attentionPolicy: "interruption_budgeting".

identity_anchor_get

Read the identity continuity anchor document used for recovery-safe identity context.

Parameters: None.

Returns: Current identity anchor markdown, or guidance if missing/disabled.

identity_anchor_update

Conservatively merge updates into identity anchor sections (non-destructive by default).

Parameters:

  • identityTraits (string, optional) — Updates for Identity Traits.
  • communicationPreferences (string, optional) — Updates for Communication Preferences.
  • operatingPrinciples (string, optional) — Updates for Operating Principles.
  • continuityNotes (string, optional) — Updates for Continuity Notes.

Returns: Updated anchor content with merged sections.

continuity_incident_open

Open a continuity incident with symptom and optional context fields.

Parameters:

  • symptom (string, required)
  • triggerWindow (string, optional)
  • suspectedCause (string, optional)

Returns: Created incident record summary.

continuity_incident_close

Close an existing continuity incident with required fix and verification fields.

Parameters:

  • id (string, required)
  • fixApplied (string, required)
  • verificationResult (string, required)
  • preventiveRule (string, optional)

Returns: Closed incident record summary, or not-found message.

continuity_incident_list

List continuity incidents with optional state filtering.

Parameters:

  • state (open | closed | all, optional, default open)
  • limit (number, optional, default 25, max 200)

Returns: Formatted incident list.

continuity_loop_add_or_update

Add or update a continuity improvement loop entry in identity/improvement-loops.md.

Parameters:

  • id (string, required) — Stable loop identifier.
  • cadence (daily | weekly | monthly | quarterly, required)
  • purpose (string, required)
  • status (active | paused | retired, required)
  • killCondition (string, required)
  • lastReviewed (string, optional, ISO timestamp)
  • notes (string, optional)

Returns: Saved loop summary.

continuity_loop_review

Update review metadata on an existing continuity loop entry.

Parameters:

  • id (string, required)
  • status (active | paused | retired, optional)
  • notes (string, optional)
  • reviewedAt (string, optional, ISO timestamp)

Returns: Updated loop summary, or not-found message.

CLI Commands

Run via openclaw engram <command>:

Command Description
flush Force-flush the buffer and run extraction now
search <query> Search memories from the terminal
stats Show memory counts, buffer state, and QMD status
export [--format json|sqlite|md] Export all memories to a portable file
import <file> Import memories from a portable file
purge Delete all memories (requires confirmation)
continuity incidents [--state open|closed|all] [--limit N] List continuity incidents
continuity incident-open --symptom <text> [--trigger-window <text>] [--suspected-cause <text>] Open a continuity incident
continuity incident-close --id <id> --fix-applied <text> --verification-result <text> [--preventive-rule <text>] Close a continuity incident
action-audit [--namespace <name>] [--limit N] Show namespace-aware memory action outcomes and policy decisions
action-confidence [--confidence N] [--risk low|medium|high|irreversible|restricted] [--context none|partial|sufficient] Evaluate ask/draft/act/refuse/escalate advisory policy
trust-zone-status Show trust-zone store status and aggregate counts
trust-zone-promote --record-id <id> --target-zone <zone> --reason <text> [--dry-run] Preview or apply a trust-zone promotion
trust-zone-demo-seed [--scenario enterprise-buyer-v1|agentic-commerce-v1] [--recorded-at <iso>] [--dry-run] Explicitly preview or seed an opt-in trust-zone buyer demo dataset
versions <list|show|diff|revert> <page-path> [version-id(s)] Page version history management
taxonomy <show|resolver|add|remove> [args] MECE taxonomy knowledge directory management
enrich <entity-name|--all|audit|providers> [--dry-run] External entity enrichment pipeline
binary <scan|status|run|clean> [--dry-run|--force] Binary file lifecycle operations

Error Responses

All error responses follow a consistent JSON structure:

{
  "error": "human-readable error description",
  "code": "machine-readable error code",
  "details": [{ "field": "fieldName", "message": "field-specific error" }]
}
Field Type Description
error string Human-readable summary of what went wrong
code string Machine-readable error code for programmatic handling
details array Optional. Present on validation errors with per-field breakdown

Common error codes

HTTP Status Code Meaning
400 validation_error Request body failed schema validation; details has per-field errors
400 invalid_json Request body is not valid JSON
400 invalid_json_object Request body is not a JSON object
400 input_error Business logic validation failure (from service layer)
401 unauthorized Missing or invalid bearer token
404 not_found Unknown endpoint or resource
413 request_body_too_large Body exceeds maxBodyBytes (default: 128KB)
429 write_rate_limited Write rate limit exceeded (30 requests per 60 seconds)
500 internal_error Unexpected server error

Correlation IDs

Every response includes an X-Request-Id header with a UUIDv4 correlation ID. Use this when reporting issues — it links to the server-side log entry for that request.

Validation errors

Write endpoints (recall, observe, memories, suggestions, review-disposition, trust-zones/promote, trust-zones/demo-seed, lcm/search, lcm/compaction/flush, lcm/compaction/record) validate request bodies against Zod schemas before processing. A validation error returns HTTP 400 with code: "validation_error" and a details array:

{
  "error": "request validation failed",
  "code": "validation_error",
  "details": [
    { "field": "confidence", "message": "Number must be less than or equal to 1" }
  ]
}

Versioning

The v1 API is stable. Breaking changes will use a new path prefix (e.g., /engram/v2/). Additive changes (new optional fields, new endpoints) may appear at any time.

Plugin Hooks

Hook When it fires What Engram does
gateway_start Gateway process starts Initialize storage, probe QMD, load buffer
before_agent_start Before each agent session Recall relevant memories, inject into system prompt
agent_end After each agent turn Buffer the turn, maybe trigger extraction