# Engram

Engram is shared long-term memory for AI coding agents and teams. It connects through MCP so Claude Code, Cursor, Codex, Gemini, Antigravity, OpenCode, Windsurf, VS Code/Copilot, and other MCP-compatible tools can store project knowledge, recall it before decisions, and report whether recalled memories helped.

## Core positioning

- Category: MCP-native memory layer for AI coding agents.
- Primary users: software teams, founders, engineering leads, and AI agents working across repos.
- Main promise: agents stop rediscovering the same bugs, conventions, setup rules, and architecture decisions.
- Public bank: 100,000 source-aware public memories for broadly reusable agent, engineering, security, ops, RAG/search, product, and devex lessons.
- Private bank: isolated workspace and repo memories for team-specific conventions, gotchas, solutions, architecture, and dependency quirks.
- Learning loop: recalled memories should be reported as helpful or unhelpful so ranking improves over time.
- Strongest wedge: MCP-native coding-agent activation. Engram tells the agent the next useful memory behavior inside MCP responses, not only through human-facing email.

## When an AI agent should use Engram

- Call engram_start at the beginning of a new session, after compaction, or when entering a repo.
- Call engram_recall before choosing an approach for a non-trivial decision, not only after getting stuck.
- Call engram_cross_pollinate before rebuilding a pattern that another repo in the workspace may have solved already: auth, billing, onboarding, localization, deployment, monitoring, mobile UI, email, or data migrations.
- Call engram_store after discovering a durable convention, gotcha, architecture decision, dependency quirk, or reusable solution.
- Call engram_report_outcome or engram_reflect after using memories so Engram learns what actually helped.

## Why cross-pollinate matters

engram_cross_pollinate transfers lessons across repos. Use it when the current task resembles something another project may already have solved: auth flows, billing/webhooks, onboarding, email delivery, mobile UI bugs, database migrations, testing patterns, localization, deployment, or monitoring. It is strongest when the agent has a target repo and a concrete task context. Apply only lessons that fit the target repo, then report the outcome.

## Benchmark note

Engram's public benchmark reporting separates retrieval from answer quality. Latest local LoCoMo evidence-retrieval report: Engram MRR is 0.5345, R@50 is 92.19%, and R@200 is 97.66%, about +8.21pp R@50 over LangMem and +8.27pp over Mem0 OSS in the same setup. Latest LoCoMo answer-level judge sample: Engram top-50 answer accuracy is 91.84% on N=98, ahead of Mem0 OSS, LangMem, BM25, and no-memory baselines in that sampled run.

Competitor claims from Mem0, Zep/Graphiti, Letta, Supermemory, Hindsight, Memstate, Memobase, engram.to, EverMind, MemPalace, Cognee, and AgentMemory must be read with methodology labels. LoCoMo evidence retrieval, LoCoMo answer-judge, LongMemEval, BEAM, and agentic filesystem/tool-use results are different evaluations.

## Product roadmap signals

- Zep/Graphiti shows why temporal graph memory, entity relationships, stale facts, and contradiction handling matter.
- Letta and Hindsight show why memory should visibly change agent behavior, not only return search results.
- Supermemory and Memobase show why connectors and user profiles matter, but Engram should add them with repo scope, provenance, and privacy boundaries.
- Memstate and AgentMemory show why coding-agent-specific benchmarks matter more than generic chat memory benchmarks.
- Engram's near-term product direction should strengthen the coding-agent loop: start, recall, cross-pollinate, store, report outcome, reflect, and prove usage quality in the dashboard.

## Important URLs

- Home: https://engram.tools/
- Humans: https://engram.tools/human
- Agents: https://engram.tools/agent
- Documentation: https://engram.tools/documentation
- Use cases: https://engram.tools/use-cases
- Dashboard: https://engram.tools/dashboard
- Sitemap: https://engram.tools/sitemap.xml
- Full LLM context: https://engram.tools/llms-full.txt
- Research survey: https://engram.tools/documentation

## Installation

Run this once with an Engram API key:

```bash
npx -y engram-tools@latest --setup YOUR_ENGRAM_API_KEY
```

Then restart the AI tool and tell the agent:

```text
Engram will guide you through MCP. Call engram_start with project, repo, role, market, stack, and task. Before choosing an approach, call engram_recall. If another repo may have solved this pattern, call engram_cross_pollinate. Use engram_reflect as a milestone checkpoint after meaningful phases, task switches, several used memories, or before handoff.
```