Skip to content

#1 Persistent memory for AI coding agents based on real-world benchmarks

docs: drop AAIF-PROPOSAL-DRAFT.md (submission paperwork, not project material)

Rohit Ghumareb995983e5a7c
diff --git a/.github/AAIF-PROPOSAL-DRAFT.md b/.github/AAIF-PROPOSAL-DRAFT.md
deleted file mode 100644
index 01bb39b9..00000000
--- a/.github/AAIF-PROPOSAL-DRAFT.md
+++ /dev/null
@@ -1,116 +0,0 @@
-# AAIF Project Proposal — draft for agentmemory
-
-Paste the fields below into <https://github.com/aaif/project-proposals/issues/new?template=project-proposal.yml> when ready.
-
----
-
-**Project Name**
-agentmemory
-
-**Project Description**
-agentmemory is a persistent, local-first memory runtime for AI coding agents. It captures session observations via 12 auto-hooks, compresses them into semantic memories, and serves retrieval through a hybrid triple-stream (BM25 + vector + temporal knowledge graph) pipeline — with zero external databases. It ships as three npm packages: `@agentmemory/agentmemory` (the server), `@agentmemory/mcp` (a standalone MCP server), and `@agentmemory/fs-watcher` (a filesystem connector). It has first-party integrations for Claude Code, OpenClaw, Hermes, and Codex CLI, and is MCP-native so any Model Context Protocol client — Cursor, Claude Desktop, Gemini CLI, VS Code, Cline, Windsurf, OpenCode, Goose, Aider — works out of the box. The project was started in January 2026 and is currently at v0.9.0 with 1,786 GitHub stars, 173 forks, and 6 contributors.
-
-**Alignment with AAIF Mission**
-agentmemory advances "the public interest in agentic AI innovation" by removing the single biggest failure mode of today's coding agents: loss of context between sessions. It does so under Apache-2.0, without vendor lock-in, and without requiring developers to stand up Redis, Postgres, Qdrant, or Neo4j. All memory is stored locally on the user's machine by default. The project is a direct, neutral consumer of the AAIF-hosted Model Context Protocol — every agentmemory tool is available through the MCP surface — which is the exact category of foundational interoperability AAIF exists to steward.
-
-**Relation to Existing AAIF Projects**
-- **Model Context Protocol (MCP)** — agentmemory is a first-class MCP server. The `@agentmemory/mcp` package is a standalone stdio MCP server exposing 44 tools; the main package registers the same tools via in-process MCP. Both pass validation against the MCP reference test suite.
-- **goose** — agentmemory works with goose via the same MCP configuration pattern documented for Claude Desktop and Cursor. We plan to ship a goose-specific integration guide in Q3 2026.
-- **AGENTS.md** — agentmemory reads `AGENTS.md` at session start and treats its contents as high-confidence context, surfaced ahead of observation-derived memory in the recall pipeline.
-
-**Example Use Cases and Evidence of Adoption**
-Core use cases:
-1. Cross-session context preservation — "what did I change in the retry logic last Tuesday?" surfaces commit, session, and file-level memories in a single recall call.
-2. Team knowledge sharing — the mesh-sync feature pushes curated memories to peer nodes over authenticated HTTPS, enabling shared memory across a dev team.
-3. Defect debugging — the session replay feature rehydrates a Claude Code JSONL transcript into full observation form, so a maintainer can re-enter the session that produced a bug.
-4. Compliance trails — every structural deletion emits an audit row (scoped for user-initiated actions, batched for automatic sweeps) per the policy in `src/functions/audit.ts`.
-
-Adoption evidence:
-- 1,786 GitHub stars, 173 forks, 6 commit contributors as of April 2026.
-- Three npm packages published with provenance and downloaded regularly (`npm view @agentmemory/agentmemory`).
-- Featured in the Claude Code plugin marketplace as `agentmemory@agentmemory`.
-- The [design doc gist](https://gist.github.com/rohitg00/2067ab416f7bbe447c1977edaaa681e2) that seeded the project has 775 stars and 106 forks.
-- 95.2% retrieval R@5 on the public LongMemEval-S benchmark, independently reproducible via the in-repo harness.
-
-Production deployments in named organizations: we are compiling case studies now and will update this section as deployers opt in to being listed. The current installation base is primarily individual developers; enterprise case studies are a Q3 2026 roadmap item.
-
-**Technical Committee Sponsor (if identified)**
-Not yet identified. We would welcome an introduction.
-
-**GitHub Repository URL**
-<https://github.com/rohitg00/agentmemory>
-
-**License**
-Apache-2.0 (OSI-approved, permissive). <https://github.com/rohitg00/agentmemory/blob/main/LICENSE>
-
-**Governance Model**
-Documented in <https://github.com/rohitg00/agentmemory/blob/main/GOVERNANCE.md>. The model follows the Linux Foundation Minimum Viable Governance pattern, scoped to the project's current single-maintainer reality with a concrete plan (see `ROADMAP.md`) to add at least one additional Maintainer from a different organization during the current growth cycle.
-
-**CI/CD & Release Workflow**
-GitHub Actions:
-- `.github/workflows/ci.yml` — runs TypeScript build + 777-test Vitest suite on every PR against Node 20 and Node 22.
-- `.github/workflows/publish.yml` — triggers on GitHub Releases. Publishes `@agentmemory/agentmemory`, `@agentmemory/mcp`, and `@agentmemory/fs-watcher` to npm with provenance, in order, with registry-propagation wait gates between them. Also exposes `workflow_dispatch` for manual publishes.
-
-Release cadence: minor releases every 2–4 weeks; patch releases as needed. Each release goes through a `release/vX.Y.Z` PR that bumps 8 in-repo version strings in lockstep (see `CONTRIBUTING.md#release-process`), is tagged `vX.Y.Z` on merge, and auto-publishes via the workflow.
-
-**Public-Facing Contribution Process for Specifications**
-<https://github.com/rohitg00/agentmemory/blob/main/CONTRIBUTING.md> documents the full PR flow, DCO sign-off requirement, coding style, subsystem layout, and the process for adding new MCP tools and auto-hooks. Non-code / policy changes follow the process in `GOVERNANCE.md#non-pr-decisions` — a `governance`-labeled issue with a 7-day public comment window.
-
-**Publicly Accessible Issue Tracker**
-<https://github.com/rohitg00/agentmemory/issues> — no gating, standard labels, issue template for bug reports in `.github/ISSUE_TEMPLATE/`.
-
-**External Project Dependencies**
-Runtime:
-- `iii-sdk` (Apache-2.0) — three-primitive engine runtime.
-- `@anthropic-ai/claude-agent-sdk` (MIT) — default LLM provider when a Claude subscription is detected.
-- `@modelcontextprotocol/sdk` (MIT) — MCP server runtime.
-- Optional: `@google/generative-ai` (Apache-2.0), `openai` (Apache-2.0), upstream minimax / openrouter — each behind a provider flag.
-
-Build-time only:
-- `tsdown` (MIT), `vitest` (MIT), `typescript` (Apache-2.0).
-
-Full lockfile at <https://github.com/rohitg00/agentmemory/blob/main/package-lock.json>.
-
-**Maintainers & Contributors**
-Current Maintainers (from `MAINTAINERS.md`):
-- Rohit Ghumare ([@rohitg00](https://github.com/rohitg00)) — Independent — project lead, all subsystems.
-
-Named non-Maintainer contributors with merged PRs:
-- [@Tanmay-008](https://github.com/Tanmay-008) — retention scoring, semantic eviction routing, multimodal image memory.
-- [@JasonLandbridge](https://github.com/JasonLandbridge)
-- [@Getty](https://github.com/Getty)
-- [@eng-pf](https://github.com/eng-pf) — security fixes.
-- [@Garygaoxiang](https://github.com/Garygaoxiang)
-
-Full contributor graph: <https://github.com/rohitg00/agentmemory/graphs/contributors>. We transparently disclose that maintainership is currently single-organization; the Q2–Q3 2026 roadmap commits to changing this.
-
-**Leadership Team & Decision Process**
-See `GOVERNANCE.md#decision-making`. Lazy consensus on PRs, formal votes on governance items (simple majority of Maintainers with a 7-day public comment window during single-maintainer periods).
-
-**Roadmap**
-<https://github.com/rohitg00/agentmemory/blob/main/ROADMAP.md>. 12 months of planned work broken into four quarterly themes: Depth (Q2), Breadth (Q3), Trust (Q4), v1.0 (Q1 2027).
-
-**Security**
-- Security policy at <https://github.com/rohitg00/agentmemory/blob/main/SECURITY.md>.
-- Six CVE-tracked v0.8.2 advisories published through GitHub Security Advisories, drafts archived under `.github/security-advisories/`.
-- OpenSSF Best Practices Silver-or-better badge: not yet enrolled. Enrollment is a Q3 2026 roadmap item.
-
-**Website URL**
-Ships in-repo under [`website/`](https://github.com/rohitg00/agentmemory/tree/main/website) (Next.js 16 App Router, deployable to Vercel with Root Directory = `website/`). Primary documentation remains the repository README.
-
-**Documented Governance Practices (optional)**
-See `GOVERNANCE.md` above.
-
-**Links to Social Media Accounts (optional)**
-- Author X: <https://x.com/ghumare64>
-
-**Details of Existing Financial Sponsorship (optional)**
-None. The project is developed in-kind by the Maintainer.
-
-**Infrastructure Needs or Requests (optional)**
-- Continued GitHub Actions CI minutes (currently within the free tier for public repos).
-- LF-hosted domain delegation if/when we move from `agentmemory.dev` to a foundation-owned hostname.
-- Potential CI time for the LongMemEval-S benchmark harness (Q2 2026 roadmap item) — the harness pulls a ~200 MB dataset and runs under ~15 minutes per release.
-
-**Additional Information**
-Applying for **Growth Stage**. We do not currently meet the Impact Stage criterion of "a diverse group of maintainers, with multiple organizations represented"; the `ROADMAP.md` commits to addressing that during the Growth cycle.