Zekra

ذكرى — shared long-term memory for AI agents. What one session learns, the next one already knows.

Zekra is a memory service that any agent (Claude Code, Codex, Gemini, Cursor, your own fleet) reaches over MCP. It stores what agents and people learn, deduplicates and updates it on write, and returns exactly the right memory on recall — with citations — so nobody re-derives what is already known.

• Landing: https://zekra.dev
• Console + API + MCP: https://app.zekra.dev
• CLI: fadymondy/zekra-cli

Documentation

Full docs at https://zekra.dev/en/docs (source: docs/):

• Introduction and Quickstart
• MCP server: tool reference and client setup
• CLI: the zekra command and the Claude Code plugin
• REST API: endpoints, errors, events, TypeScript client
• Data sources · Capture hook · Security · Self-hosting

What it does

• Retain / recall — hybrid retrieval: dense vectors + multilingual BM25 fused with RRF, a salience nudge, reranking, and 1-hop entity expansion. A write-decision step (ADD / UPDATE / INVALIDATE / NOOP) updates contradicted facts instead of duplicating them. Nothing is hard-deleted: forget soft-invalidates and history stays queryable.
• Many brains — one namespace per project or team, with per-agent grants and access tokens (cbt_…) enforced server-side.
• Knowledge graph — entities and typed relations extracted from memories, with traverse / neighbors / path / ontology queries and a graph "spine".
• Data sources — connectors (text, markdown, crawler, github, sql, webhook) pull external content, chunk it and retain it through the same write pipeline. See Data sources.
• Chat with a brain — a tool-calling agent that uses the brain's own recall / search / graph / retain as tools and answers with citations and an auditable trace.
• Knowledge gaps, secrets vault, activity log — plus a web console to manage it all.
• Tiers — Redis L1 working-memory cache, hot Postgres tier, object-store cold tier.

Connect an agent

curl -fsSL https://app.zekra.dev/install.sh | sh     # installs the `zekra` CLI
zekra auth login --token <cbt_…>
zekra mcp:install claude-code                        # or claude-desktop | codex | gemini | cursor

MCP tools include memory_retain, memory_recall, memory_get, memory_forget, memory_share, memory_edit, graph_*, memory_gaps, brain_* (list, grants, tokens, chat), secret_* and datasource_*. Full reference: MCP server · contracts: contracts/tools.md.

Architecture

Zekra is a togo app built from plugins:

Storage is Postgres with VectorChord + BM25 + pgvector. Design: SPEC.md · build plan: PLAN.md · decisions: contracts/internal/decisions.md · deploy: DEPLOY.md.

Develop

cp .env.example .env
togo generate        # sqlc → gqlgen → atlas → OpenAPI (the build gate)
togo migrate         # harness schema;  go run ./cmd/zekractl migrate  for the brain schema
togo serve           # API + console
go test ./plugins/brain/...
cd web && npm run dev

Configuration comes from .env / togo.yaml (ZEKRA_* variables; legacy CABRAIN_* names are still honoured). Secrets are never committed.

License

Zekra is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0-only). You can use, modify and self-host it; if you run a modified version as a network service, you must offer its users the corresponding source under the same license.