It is not a vector database (vectors are one retrieval mode among six). It
is not just a memory store (knowledge is source-grounded, not
free-floating; memory is experiential, not factual). It is the layer where raw
observations become durable beliefs, where code becomes a queryable
graph, and where policy, provenance, and scope govern every read and
write — so both knowledge and memory are auditable, not accidental.
Local-first agentic memory: the engram MCP server launcher. Memory + knowledge graph + beliefs + code intelligence for any MCP client.
2026-11-15
스타
—
도구
—
가동 시간
—
Install
Quick start (any MCP host)
# Install the server (alpha)
npm install -g @engramlabs/mcp@alpha
# Register with your agent — one command, no JSON editing
engram-mcp install --client claude # Claude Code (project scope)
engram-mcp install --client claude-desktop # Claude Desktop (user scope)
engram-mcp install --client cursor # Cursor (project scope)
engram-mcp install --client codex # Codex CLI (user scope)
engram-mcp install --client pi # pi (project scope)
engram-mcp install --client copilot # GitHub Copilot / VS Code
# Or run directly without installing
npx @engramlabs/mcp@alpha --storage ~/.engram
Agent bundle — skills, hooks, and an adversarial subagent (one command)
engram-mcp install --agent claude # or codex | copilot | pi | all
engram-mcp install --agent claude --remove # uninstall (name-tagged: only engram-owned entries go)
What lands on your machine:
Piece
What it does
7 skills
engram-index (index + memorize with evidence labels), engram-know (read-only evidence-chained query + verify), engram-mem (decisions, gotchas, procedures, stances), engram-distill (source-keyed knowledge-graph distillation), outcome-reporting (report whether recall actually helped), engram-factory (spec-driven software factory), engram-frame (intent framing before a factory job). The set derives from the repo's authored skill tree — new skills ship automatically. Retired skills from older installs (engram-search, engram-plan, engram-implement, engram-learn) are removed automatically unless you replaced them
Factory specialists
engram-factory ships its 14 specialist subagents (analyzer, planner, tasks, builder, eight discipline reviewers, qa, approver) in each host's native dialect — claude, codex, copilot, pi
Recall hooks
Every prompt recalls relevant memory — gated: irrelevant results stay silent (separation gate, no LLM in the hot path). Session-start packs, stale-index reminders, kill switch (ENGRAM_HOOKS=0), fail-open always
red-forman subagent
Read-only adversarial reviewer — severity-tagged findings (BLOCKER/DEFECT/SMELL) with file:line evidence and a one-line fix each. Dispatch it before declaring work done
Per-platform notes: Codex agents use YAML (~/.codex/agents/*.yaml); Copilot command hooks cannot inject per-prompt context (platform limitation — session-start recall only, honestly documented); Copilot agents are repo-scoped either way — even --scope user writes 15 agent files into the current repo's .github/agents/, so run that install from the repo that should carry them; pi gets an in-process extension; claude-desktop hosts MCP registration only (no skills/agents filesystem — install --agent all prints that limitation).
Project scope — install the bundle into the repo instead (committed, shared with the team):
cd your-repo
engram-mcp install --agent all --scope project
What you get
Your agent gains these MCP tools:
Category
Tools
What they do
Discover
recall, search, get_context, explore
Find anything across memory, knowledge, code — one query, fused results
Track derived stances, open tensions, replayable runbooks
Content
fetch_content
Stitch any indexed file's full content back — docs or code
Correct
forget, graph_health
Remove wrong data, monitor store health
Platform support
OS
Architecture
Status
Linux
x86-64 (glibc)
✅
Linux
ARM64 (glibc)
✅
macOS
Apple Silicon (M1–M4)
✅
Windows
x86-64
✅
Privacy: engram is local-first. All data lives in SQLite files on your
machine. No telemetry, no cloud calls, no network on the default path —
the code has no network usage outside optional HTTP embeddings you
configure yourself.
For product, data, and strategy teams — the problems engram solves, in plain terms:
Long-horizon agent memory. An agent that remembers decisions, constraints,
and stakeholder preferences across sessions, days, or months — without
re-reading its entire history every turn. Less context burned, better answers
on multi-hop questions.
Source-grounded knowledge, not hallucinated facts. Ask questions grounded
in your code repositories and documents; every answer traces back to a chunk
and its source. Memory (agent experience) stays distinct from knowledge
(grounded content) so the two never blur.
Auditable, governed recall. Every record carries policy, provenance, and
scope — who saw it, where it came from, how long it lives, who may retrieve
it. Redaction and deletion are first-class; nothing leaks through retrieval.
Code understanding. A codegraph layer answers structural questions — "who
breaks if I change this?", "find the call path from handler to database",
"what changed recently that matters most?" — over any indexed repository.
Concepts at a glance
Engram bakes in eight concepts beyond "store text, search by similarity." The
full version is in The conceptual model; the pipeline
that connects them is in the architecture overview.
Memory as a first-class domain — typed records with a lifecycle, not bare rows.
Source-grounded knowledge graph — facts traceable to a source, never free-floating.
Belief synthesis — derived, recomputable, bi-temporal stances over evidence.
Hierarchy — aggregate + navigate + compress context to the right granularity.
Retrieval composition — six modes fused (RRF) + cross-encoder rerank, never one.
Consolidation — reflection + decay as an explicit, auditable pipeline.
Policy, provenance, scope — governance checked on every read and write.
Contract-first — domain types outrank SQL; portable across storage engines.
Why engram exists
Agent memory gets messy when storage, ranking, policy, provenance, and runtime
integration collapse into one service. The result is either a vector DB that
returns similar text without understanding why it's relevant, or an LLM
context window that forgets everything when the session ends.
Engram keeps these concerns separate and explicit:
Contract-first domain types that rank above any SQL schema or storage
engine — so agents depend on semantics, not implementation.
Rust-owns-deterministic-behavior — the core traits, adapters, and
conformance tests are in Rust; TypeScript wraps generated contracts.
Replaceable adapters — storage (SQLite, Postgres/pgvector), retrieval (vector,
graph, lexical, associative), and embedding providers (FastEmbed, Ollama) are
behind ports, never hardcoded.
Policy on every path — scope, retention, allowed-uses, and provenance are
checked on write, retrieve, ingest, consolidate, and forget — never hidden in
a generic manager.
The conceptual model
Engram bakes in eight concepts that go beyond "store text, search by
similarity." Each is a distinct domain axis with its own port, lifecycle, and
contract.
1. Memory as a first-class domain (not just rows)
A MemoryRecord is a canonical durable unit with explicit kind
(observation, fact, preference, episode, artifact, relationship, procedure),
content, scope, provenance, policy, status, and links — never a
bare row. Memory has a lifecycle (active → archived → redacted → forgotten → expired) and emits append-only MemoryEvents (written, retrieved,
consolidated, forgotten, belief_synthesized, …). State and events are
separated, so the layer is auditable and replayable rather than a mutable
log.
A MemoryRole (working / episodic / semantic / procedural, CoALA-aligned) is
derived from kind + policy + scope + provenance — keeping the contract small
while preserving the cognitive-science taxonomy.
Knowledge is a separate domain axis from memory:
KnowledgeSource → SourceDocument → KnowledgeChunk → KnowledgeEntity / KnowledgeRelationship → EmbeddingRef, bounded by a named KnowledgeGraph.
Memory records agent experience; knowledge records are source-grounded
content from code repos, documents, URLs. A memory may link to a knowledge
chunk but cannot replace it — preventing hallucinated "facts" from masquerading
as grounded knowledge. Every chunk references its source and document (invariant).
3. Belief synthesis (what agents BELIEVE vs what they observed)
A Belief is a derived stance over evidence — never raw memory, never source
truth. It has a subject, declarative content, confidence, weighted
BeliefSources, a lifecycle (active / stale / superseded / retracted), and a
synthesizer derivation ref. A belief is recomputable — when a source is
invalidated, a single-source belief retracts; a multi-source belief is marked
stale and resynthesized.
Beliefs are bi-temporal: valid_from / valid_until define when the content
was true in the world (valid time), distinct from created_at / updated_at
(when engram recorded it — transaction time). This "true as-of T" retrieval is
the whitespace no competitor owns.
Contradiction records are reviewable signals, never automatic truth changes
— they classify the tension (logical, temporal, tension, duplicate,
policy) and capture how a human or system resolved it.
HierarchyNode organizes retrievable objects (memories, chunks, entities,
concepts, other nodes) into abstraction layers — layer 0 = base retrievable,
layer > 0 = aggregate (summary, schema, topic, cluster, domain). This is the
GraphRAG insight: a query may need a raw chunk, an episode summary, a workflow
schema, or a domain pattern; hierarchy lets retrieval return the right
granularity instead of N similar low-level fragments.
Hierarchy ≠ taxonomy. Taxonomy organizes controlled concepts; hierarchy
organizes retrievable objects for navigation and context compression.
Construction (building the tree) is separated from navigation (traversing it).
5. Retrieval composition (never one mode)
RetrievalRequest fans out across six modes — temporal, cue,
hierarchical, semantic (vector), graph, keyword (BM25) — with results
fused, reranked, budget-compressed, and policy-filtered into a ContextPayload.
The RetrievalScore is multi-factor: relevance + recency + confidence + cue_match + hierarchical_fit + policy_fit. Fusion uses
Reciprocal Rank Fusion
(RRF) to unify heterogeneous sources without distributed write consistency.
An optional cross-encoder reranker reorders fused candidates for precision.
A FusionTrace is attached to every result: source, source rank, source score,
fusion score, rerank score, deduplicated-with — full explainability of how a
candidate moved through the pipeline. When sources degrade or results are
omitted (budget, policy), the agent is told what was dropped and why.
6. Consolidation (reflection + decay, as an explicit pipeline)
Consolidation is a first-class operation pipeline, not an incidental write
side-effect: ConsolidationRequest → ConsolidationPlan (dry-run) → ConsolidationRun (auditable execution) → ConsolidationTaskResults.
This is where reflection happens (working memory → episodic event → semantic
fact → optional taxonomy/graph/procedural update — Generative Agents /
Reflexion-style) and where forgetting happens (Ebbinghaus-style decay:
policy-expiry + memory curve). Both are explicit, both auditable. Without
consolidation, memory degrades into an undifferentiated log.
Task kinds: fact_extraction, belief_synthesis, belief_contradiction_detection,
hierarchy_build, taxonomy_evolution, graph_evolution, decay, pruning,
procedure_extraction, evaluation, and more.
7. Policy, provenance, and scope (governance baked into every path)
Policy (visibility / retention / sensitivity / allowed_uses / expires_at / delete_mode) and Provenance (source / actor / observed_at / evidence / derivations / confidence / method) are required fields on every durable
record. Scope (tenant required; subject / workspace / session / environment optional) bounds every operation.
These are not optional metadata — they are checked at runtime on write,
retrieve, ingest, consolidate, and forget. Redacted records must not leak
through retrieval, evaluation, links, or explanations. Forgetting is a domain
concept (delete / redact / tombstone / archive), not a DB delete. Every
record's lineage is traceable via EvidenceRef and DerivationRef.
The human-readable domain model (docs/domain-data-model.md), the
machine-readable wire contract (contracts/v1/), and the product overview
(docs/product/engram.md) are the source of truth. Rust types, generated TypeScript,
JSON Schema, database schemas, and API payloads all conform downward — never
the reverse. This is what makes engram portable across SQLite today and
Postgres / Neo4j tomorrow without breaking agents that depend on it.
The contract freeze policy forbids renames, removals, or meaning changes within
a version; breaking changes require a new versioned contract. metadata maps
are allowed, but core semantics must use typed fields (no smuggling policy or
provenance through unstructured metadata).
Storage backends — one crate per engine, swap by config
Each storage engine lives in a single crate holding all its database
operations. Consumers switch backends by changing a config string + Cargo feature
— no application code changes, no migration (fresh store on switch).
# Select the backend in a profile file
[backend]
kind = "sqlite" # or "postgres"
data_root = "/var/lib/engram"
# Compile with the backend feature
cargo build --features sqlite
Both backends share the same ports, same DTOs, same facade — the storage
engine is an adapter detail. An engine-neutrality lint
(.codex/hooks/check-engine-neutrality.sh) enforces that no engine type
(Sql*, raw SQL) leaks into the neutral facade or port crates.
Engine-agnostic adapters (lexical BM25, associative-graph PPR, community-summary,
cross-encoder rerank, decay, ingest) work with any backend — see
Extend the storage layer for the full
adapter list + how to add a new engine.
Architecture
Data flows one way: an agent writes (or ingests a source) → engram
processes and stores across swappable engine cells (policy + provenance on
every row) → a separate retrieval composition layer fuses many recall modes
into one policy-filtered context packet returned to the agent. Storage and
retrieval are deliberately decoupled — the same "decouple what is stored from
how it is retrieved" insight as Memora.
The rule of thumb: engram-domain owns portable concepts; engram-memory,
-knowledge, -belief, -hierarchy, -retrieval, -consolidation own their
respective ports; engram-integration is the SDK facade (EngramProvider);
concrete infrastructure lives behind adapter crates; TypeScript wraps generated
contracts instead of redefining them.
The full pipeline diagram (write/ingest → process → storage cells → retrieval
composition → context packet), the layer responsibilities, and the research
grounding are in the architecture overview.
Status
Engram is pre-1.0 — demo-driven, not production-ready. The conceptual
model (memory, knowledge, belief, hierarchy, retrieval, consolidation) is
direction-fixed; the frozen v1 contract covers the memory + retrieval +
evaluation vertical. Belief, contradiction, hierarchy, and consolidation are
draft extension contracts — direction-fixed but not frozen.
Engine-neutrality lint: enforces no engine types in the facade/ports.
Codegraph layer (RFC-0012): dead_code, blast_radius, dependency_path,
central_symbols (PageRank), bridge_symbols (betweenness),
call_communities (Louvain), temporal scoring — on top of engram, not in it.
let config = EngramConfig::from_profile_file("engram.toml")?;
let provider = EngramProvider::open(&config)?;
Connect via MCP
The current server is the unified engram-mcp (RFC-0015) — one server
covering memory + knowledge graph + beliefs + hierarchy + code intelligence.
See mcp/engram-mcp/HOW-TO.md to build + wire it.
The two-server (engram-memory-mcp / engram-codegraph-mcp) model below is
retired — those crates are deleted; the table is kept as historical reference.
Engram ships two MCP servers exposing memory + codegraph operations as
agent-callable tools (stdio JSON-RPC 2.0), so any client — Claude Desktop,
Cursor, Copilot — can read and write engram with no code on your side:
// e.g. .vscode/mcp.json or claude_desktop_config.json — stdio transport
{
"mcpServers": {
"engram-memory": {
"command": "cargo",
"args": ["run", "-p", "engram-memory-mcp", "--", "/path/to/store"]
}
}
}
Full tool lists, the codegraph index-then-query flow, per-client config
locations, and the MCP-vs-N-API-vs-Rust-facade choice are in the
MCP guide.
Contracts
The accepted v1 contract package lives in contracts/v1/. Domain types in
engram-domain are the Rust source of truth; TypeScript types are generated
from them and should not be edited by hand.
pnpm run contracts:generate # generate TS types from Rust
pnpm run contracts:check-generated # verify they match
python3 tools/scripts/validate_contracts.py
Development workflow
Engram uses spec-driven implementation:
Record durable architecture decisions in docs/adr/.
Add or update specs under docs/specs/ before behavior changes.
Run Rust, TypeScript, contract, docs, and engine-neutrality gates.
이 페이지에서 Cursor, Claude Desktop, Claude Code 또는 VS Code용 stdio 설정을 복사합니다. PlugMCP 프록시가 아니라 게시자의 설치 방식을 사용합니다.
@engramlabs/mcp에 API 키가 필요한가요?
이 레코드에는 필수 환경 변수가 등록되어 있지 않습니다.
@engramlabs/mcp은(는) 어떤 전송 방식을 사용하나요?
@engramlabs/mcp은(는) stdio를 통해 로컬에서 실행됩니다.
@engramlabs/mcp은(는) 마지막으로 언제 확인되었나요?
마지막 확인: 2026-11-15. Local-first agentic memory: the engram MCP server launcher. Memory + knowledge graph + beliefs + code intelligence for any MCP client.
