Skip to content
Engram Docs

Roadmap

Vision

Section titled “Vision”

ENGRAM is the memory substrate for AI agents: the layer that lets agents store what they learn and recall it by meaning so they get more capable over time. The roadmap is sequenced so the cognitive memory core works end-to-end before we add the human-facing dashboard and broader platform features.

Where we are

Section titled “Where we are”

The foundation is built: the NestJS MCP server, Postgres and Prisma, Postgres-backed short-term and long-term memory, an embeddings package, a pgvector vector-store package, health checks, and memory CRUD MCP tools.

The vector lifecycle is wired end-to-end: embeddings are generated, indexed in pgvector, and served through the semantic recall path.

Delivery waves

Section titled “Delivery waves”

Each wave maps to a GitHub milestone and a set of epics. Track work through the epic:* labels on each issue.

Wave Milestone Epics
1. Recall that works v0.1 — Semantic Recall Semantic Memory Engine (#98), Retrieval Quality & Evaluation (#102)
2. Memory that thinks v0.2 — Memory Intelligence Memory Intelligence (#99), Agent Ergonomics & SDK (#100)
3. Multi-tenant and portable v0.3 — Multi-tenant & SDK Auth & Multi-tenancy (#101), Agent SDK (#100)
4. Production and dashboard v1.0 — Production & Dashboard Admin API & Dashboard (#103), Production Readiness (#104)

Epic summary

Section titled “Epic summary”
Epic Focus
#98 Wire the vector lifecycle, add semantic recall, hybrid search, and ranking
#99 Consolidation, decay, deduplication, contradiction reconciliation, and insights
#100 High-level remember/recall/forget/reflect tools, scoping, and a client SDK
#101 OAuth, JWT, organizations, API keys, and strict per-tenant isolation
#102 Relevance and latency evaluation harness with CI regression gates
#103 tRPC API and web dashboard to browse, search, and monitor memories
#104 Security hardening, observability, deployment, backups, and load testing

Near-term sequence

Section titled “Near-term sequence”
  1. Define the vector-store abstraction over pgvector (#105).
  2. Wire vector upsert and delete into the long-term memory lifecycle (#107).
  3. Implement semantic kNN search and the recall MCP tool (#108, #109).
  4. Stand up the relevance evaluation harness to keep recall quality honest (#113).
  5. Add hybrid search and relevance ranking (#110, #111).

Quality gates

Section titled “Quality gates”

Run these checks before opening a pull request when the touched area supports them:

Terminal window
pnpm docs:check
pnpm build
pnpm lint
pnpm typecheck
pnpm test

For the MCP server specifically:

Terminal window
pnpm --filter mcp-server lint
pnpm --filter mcp-server test

Working agreements

Section titled “Working agreements”
  • Start from a feature branch, not main.
  • Keep changes tied to the issue or request in front of you.
  • Prefer existing workspace packages and framework CLIs.
  • Update the README and setup docs when startup commands change.
  • Keep detailed implementation notes in focused docs rather than expanding the root README.
Section titled “Links”