# Engram Full LLM Context

> Persistent, searchable memory for AI agents and teams. Engram carries durable knowledge across sessions, MCP tools, repositories, and teammates.

Last updated: 2026-10-08

## Product model

Engram is an MCP-native memory layer for AI agents. It is not a static config file and it does not replace repository inspection, logs, tests, or engineering judgment. It gives agents a durable starting map: prior decisions, conventions, production gotchas, reusable patterns, dependency quirks, and outcome-tested lessons.

- Public memory: 100,000 source-aware, reusable memories available across workspaces.
- Private memory: workspace and repository knowledge isolated to the team.
- Umbrella workspace resolution: hierarchical workspace memory resolution across related sub-repositories without duplicating knowledge.
- Memory types: CONVENTION, PATTERN, GOTCHA, SOLUTION, ARCHITECTURE, DEPENDENCY.
- Learning loop: agents report whether surfaced knowledge helped; Engram uses those outcomes for ranking and confidence.
- Sparse feedback pulse: agents can submit quality signals (useful, noisy, stale, missing_context, too_long) via engram_reflect or POST /api/v1/feedback.
- Cross-repository learning: an agent can retrieve lessons from related work while still checking whether they fit the target repository.

## Supported clients

- Claude Code
- Cursor
- Codex
- Gemini
- Antigravity
- Copilot
- OpenCode
- Windsurf
- Devin
- Cline
- Roo Code
- Kiro
- Amazon Q Developer
- Goose
- Continue

Any MCP-compatible client can connect when it supports the required transport and configuration.

## Installation boundary

A human workspace owner creates or authorizes the workspace credential. An agent with approved shell and file access can then preview and apply setup:

    export ENGRAM_API_KEY=engram_...
    npx -y engram-tools@latest --setup --client codex --dry-run
    npx -y engram-tools@latest --setup --client codex

Agents must not invent credentials, expose secrets, silently create paid plans, or grant themselves owner or billing privileges. Restart the MCP client after setup.

## Tool reference

- engram_start: Start a new session with setup guidance, relevant private and public memory, and matching skills.
- engram_rehydrate: Recover the latest session checkpoint, decisions, gotchas, and open work after context compaction or resume.
- engram_recall: Retrieve relevant private and public knowledge before a non-trivial decision.
- engram_cross_pollinate: Find transferable lessons from other repositories in the workspace before rebuilding a known pattern.
- engram_store: Store a durable convention, pattern, gotcha, solution, architecture decision, or dependency quirk.
- engram_report_outcome: Report whether recalled knowledge helped so future ranking can improve.
- engram_reflect: Create a checkpoint after a meaningful phase, before handoff, or after several memories were used. Accepts optional sparse feedback pulse (signal, notes, dismiss).
- engram_smarter: Reload focused knowledge after switching tasks or losing context mid-session.
- engram_get_context: Load a focused repository knowledge dump for advanced inspection or debugging.
- engram_skill: Load a step-by-step guide when a task needs a repeatable procedure.
- engram_proof: Return workspace memory, activation, usage, and benchmark evidence for product or administrative review.
- engram_setup: Show workspace and setup guidance; normal sessions should prefer engram_start.

## Recommended behavior

1. Call engram_start at the beginning of every meaningful session or when entering a repository.
2. Call engram_rehydrate before continuing after compaction, summary, truncation, resume, or missing context.
3. Call engram_recall before choosing an approach for a non-trivial decision, not only after getting stuck.
4. Call engram_cross_pollinate when authentication, billing, onboarding, localization, deployment, monitoring, mobile UI, email, data migration, or another reusable pattern may have been solved in a different repository.
5. Store only durable knowledge. Never store credentials, raw customer data, or temporary debug noise.
6. Verify recalled knowledge against current code and runtime state.
7. Report outcomes after using recalled memory. Use engram_reflect to checkpoint substantial work and optionally submit sparse feedback pulse.

## Benchmark evidence

Engram publishes evidence retrieval and answer quality separately because a larger retrieved set can improve recall while adding distracting context for the answering model.

Evidence retrieval, local LoCoMo diagnostic (N=1,536):

- Engram: MRR 0.5427, R@50 93.29%, R@200 97.59%.
- Mem0 OSS 2.2.0: MRR 0.4314, R@50 85.16%, R@200 94.79%.
- LangMem / LangGraph Store 0.0.30: MRR 0.4006, R@50 83.98%, R@200 94.53%.
- AgentMemory 0.9.29: MRR 0.4627, R@50 83.33%; its API caps retrieval at 100, so no R@200 is reported.
- BM25: MRR 0.4676, R@50 78.19%, R@200 85.74%.

Answer diagnostic, full shared evidence-bearing set (N=1,536):

- Engram: 1,377/1,536 correct, 89.65%.
- Mem0 OSS: 1,294/1,536 correct, 84.24%.
- AgentMemory: 1,280/1,536 correct, 83.33%.
- LangMem / LangGraph: 1,279/1,536 correct, 83.27%.
- Session-file BM25: 1,235/1,536 correct, 80.40%.
- BM25: 1,227/1,536 correct, 79.88%.

These are checked-in, locally reproduced diagnostics, not official vendor submissions or leaderboard scores. The same answer model and a separate blind judge were used for every published system. An empty-context control remains in the internal artifact to verify the evaluation pipeline, but it is not presented as a competing memory system. Read the methodology and limitations before comparing results from different systems or datasets.

## Plans and limits

- free: $0/month, 1 member, 100 calls/day, 500 private memories.
- developer: $9/month, 2 members, 1,000 calls/day, 5,000 private memories.
- team: $29/month, 7 members, 5,000 calls/day, 25,000 private memories.
- scale: $99/month, 30 members, 25,000 calls/day, 100,000 private memories.

For more than 30 members, book a call. Live pricing remains the canonical commercial source: https://engram.tools/en#pricing

## Canonical sources

- Home: https://engram.tools/en
- Human setup: https://engram.tools/en/human
- Human product deck: https://engram.tools/en/deck/human
- Agent pitch deck: https://engram.tools/en/deck/agent
- Agent guide: https://engram.tools/en/agent
- Machine-readable agent guide: https://engram.tools/agent.txt
- Documentation: https://engram.tools/en/documentation
- Benchmarks and limitations: https://engram.tools/en/benchmarks
- Real use cases: https://engram.tools/en/use-cases
- Shared memory for AI agents: https://engram.tools/en/ai-agent-memory
- MCP memory server: https://engram.tools/en/mcp-memory-server
- Context compaction recovery: https://engram.tools/en/context-compaction-recovery
- Benchmark methodology and claim rules: https://engram.tools/en/benchmark-methodology
- Articles and research: https://engram.tools/en/insights
- How to evaluate AI agent memory: https://engram.tools/en/insights/ai-agent-memory-benchmarks
- Claude Code integration: https://engram.tools/en/integrations/claude-code
- Codex integration: https://engram.tools/en/integrations/codex
- Gemini integration: https://engram.tools/en/integrations/gemini
- OpenCode integration: https://engram.tools/en/integrations/opencode
- Same-harness Mem0 comparison: https://engram.tools/en/compare/mem0
- Same-harness LangMem comparison: https://engram.tools/en/compare/langmem
- Same-harness AgentMemory retrieval comparison: https://engram.tools/en/compare/agentmemory
- Short machine-readable context: https://engram.tools/llms.txt
- Sitemap: https://engram.tools/sitemap.xml