Thred

Revision-aware, cross-session memory for AI agents.

Thred is a durable memory layer for agents working across Cursor, Claude Code, Codex, and OpenCode. It turns long, fragmented chat histories into verifiable memory — decisions, facts, preferences, source evidence, revisions, and resumable coding handoffs.

Get started → thred.fun · MIT License

The problem

Cross-session agent work breaks down when the relevant information is spread across dozens of conversations. Long-running projects can span 30–40 sessions and more than 115K tokens per question. A useful memory layer has to do more than retrieve semantically similar text:

• combine facts from many sessions while retaining chronology;
• represent facts that changed or were overwritten;
• show where an answer came from; and
• return NOT_FOUND when the answer is not supported by the history.

Copy-pasting the last session into the next prompt works for a short handoff, but not across dozens of sessions: the context window fills up, older facts get missed, and agents confuse superseded decisions with current ones. Thred fixes that with atomic, provenance-backed facts, explicit revision history, and honest abstention when the evidence isn't there.

Features

Thred exposes a small MCP toolset — use the same memory from any agent client:

Tool	What it does
thread_remember	Saves a durable, evidence-backed fact, decision, lesson, architecture choice, or preference.
thread_context	Retrieves verified long-term context relevant to a question.
thread_history	Shows the chronological revision history for a fact or entity.
thread_inspect	Traverses a memory’s provenance and graph relationships.

Save "we chose PostgreSQL" in one session. Update to "we chose Neon Postgres" in another. Ask what’s true now — or why it changed — in a third. Nothing gets silently erased.

Quick start

1. Connect an MCP client

Create an agent key at thred.fun, then add this to Cursor, Claude Code, Codex, or OpenCode:

{
  "mcpServers": {
    "thred": {
      "command": "npx",
      "args": ["-y", "@thred_nick_01/thred-mcp"],
      "env": {
        "THRED_API_KEY": "thrd_sk_…",
        "THRED_API_URL": "https://api.thred.fun"
      }
    }
  }
}

Use the same key across clients to share one workspace memory.

2. Run locally (optional)

npm install
npm run db:generate --workspace=@repo/db
npm run db:migrate --workspace=@repo/db
npm run dev

Prerequisites: Node.js 18+, PostgreSQL, HydraDB API key, and Groq or OpenAI.

Variable	Required for	Notes
DATABASE_URL	API, frontend, evaluations	PostgreSQL connection string.
HYDRA_DB_API_KEY	API, memory engine, evaluations	HydraDB access token.
OPENAI_API_KEY or GROQ_API_KEY	Extraction and evaluations	Configure at least one provider.
BETTER_AUTH_SECRET	Web authentication	Signing/encryption secret.
BETTER_AUTH_URL, FRONTEND_ORIGIN	Web authentication	Local defaults point to http://localhost:3000.
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET	Google sign-in	Required only when Google auth is enabled.
PORT, NEXT_PUBLIC_API_URL	Local development	Optional API port and frontend API URL overrides.

For MCP clients, configure THRED_API_KEY; THRED_API_URL defaults to the production API.

How it works

1. Agent sessions are processed into atomic, evidence-backed claims.
2. Claims are stored with workspace, session, source-message, entity, and file provenance.
3. When a claim changes, Thred writes an explicit SUPERSEDES link to the prior claim.
4. At query time, hybrid retrieval and graph context surface candidate evidence.
5. A temporal resolver returns current or historical facts as requested.
6. Insufficient evidence → NOT_FOUND, not a guess.

Built on HydraDB for graph-backed storage with hybrid retrieval — so revision history and provenance are first-class, not bolted onto a vector store.

Benchmarks

Evaluated against a hashed-vector RAG baseline on LongMemEval, LongMemEval V2, and BEAM.

Evaluation	Vector-RAG	Thred	What it demonstrates
LongMemEval V2 scale case	100.0%	100.0%	Correct temporal answer across 44 sessions / ~128K tokens; zero evaluation errors.

Full reports in apps/evals/reports.

npm run test --workspace=@repo/evals
npm run test --workspace=@repo/memory-extractor
npm run test --workspace=@repo/memory-engine

Tech stack

TypeScript · Node.js · Express · Next.js · React · MCP · HydraDB · Prisma/PostgreSQL · OpenAI/Groq