anda-brain

🧠 Anda Brain

Persistent long-term memory service for LLM agents, powered by a Knowledge Graph (Cognitive Nexus) and KIP (Knowledge Interaction Protocol). Anda Brain is open-source software designed to be self-hosted — deploy your own instance with the Quick Start guide.

Note: The hosted cloud service (brain.anda.ai) and its console (anda.ai/brain) have been discontinued. All examples below assume your own deployment.

For a complete, ready-to-run agent built on Anda Brain, see Anda Bot.

Business agents interact entirely through natural language and a simple REST API — no KIP knowledge required.

Business Agent  ──natural language──▶  Brain  ──KIP──▶  Cognitive Nexus
 (your agent)                         (this service)          (knowledge graph)

What You Get

Three operational modes cover the full memory lifecycle:

Supporting endpoints:

Auth scopes in tables apply when authentication is enabled (ED25519_PUBKEYS is set).

When to Use This Service

Use Anda Brain when your agent needs to:

• Persist knowledge across sessions — user preferences, facts, decisions, relationships, events
• Recall previous context — what happened before, what the user said, what decisions were made
• Share memory across agents — multiple agents can read/write to the same space
• Maintain memory health — consolidate old events, deduplicate facts, decay stale knowledge

The service handles all the complexity of knowledge graph management. Your agent just sends messages and asks questions in natural language.

When NOT to Use

• Temporary conversation context that only matters in the current session
• Large file storage (use object storage instead)
• Real-time data streaming
• Secrets, passwords, or API keys (the service is not a vault)

Concepts

Memory Space

Each space is an isolated environment with its own knowledge graph, conversation history, and database. Spaces are identified by a space_id string.

Memory Types

The Formation agent extracts three types of memory from conversations:

1. Episodic Memory (Events) — What happened, when, who participated, outcome
2. Semantic Memory (Stable Knowledge) — Facts, preferences, relationships, domain knowledge
3. Cognitive Memory (Patterns) — Behavioral patterns, decision criteria, communication style

Cognitive Nexus

The underlying knowledge graph consists of:

• Concept Nodes — Entities with a type and name (e.g., {type: "Person", name: "Alice"}, {type: "Preference", name: "dark_mode"})
• Proposition Links — Directed relationships between concepts (e.g., (Alice, "prefers", dark_mode))

Authentication

If ED25519_PUBKEYS is configured, protected endpoints require a Bearer token in the Authorization header.

If ED25519_PUBKEYS is empty/not provided, authentication is disabled and requests are accepted without signature verification.

Authorization: Bearer <base64_encoded_cose_sign1_token>

Management endpoints (/v1/{space_id}/management/*) and admin endpoints (/admin/*) still follow their role/scope checks when auth is enabled.

API Reference

For complete endpoint and TypeScript schema details, see:

• https://github.com/ldclabs/anda-brain/blob/main/anda_brain/API.md (English)
• https://github.com/ldclabs/anda-brain/blob/main/anda_brain/API_cn.md (中文)

Content Negotiation

The API supports triple serialization. Set Content-Type and Accept headers accordingly:

• application/json — JSON (default)
• application/cbor — CBOR (binary, more compact)
• text/markdown — Markdown (raw text or formatted Markdown)

All responses are wrapped in an RPC envelope when using JSON or CBOR:

{
  "result": { ... },
  "error": null
}

When Accept: text/markdown is used, the response is returned as raw text or a Markdown formatted string.

On error:

{
  "result": null,
  "error": {
    "message": "error description",
    "data": { ... }
  }
}

Markdown Serialization Sample

If Accept: text/markdown is specified, the result field's content will be directly serialized as the response body.

Request:

POST /v1/my_space_001/recall
Accept: text/markdown

What are Alice's preferences?

Response (HTTP 200):

Alice has the following known preferences:
- **Dark mode** in all applications (confidence: 0.9, since 2025-01-15)
- **Email communication** preferred over phone calls (confidence: 0.8, since 2025-01-10)

Alice is currently working on **Project Aurora** and was last seen on 2025-01-15 discussing settings preferences.

Gaps:
- No information found about Alice's language preferences.

Create Space

Create a new isolated memory space.

POST /admin/create_space
Authorization: Bearer <token>
Content-Type: application/json

Request:

{
  "user": "<owner_principal_id>",
  "space_id": "my_space_001",
  "tier": 0
}

Response:

{
  "result": { ... }
}

Formation — Encode Conversations into Memory

Send conversation messages to be analyzed and encoded into the knowledge graph. The service extracts facts, preferences, relationships, events, and patterns, then stores them as structured knowledge.

Processing is asynchronous — the endpoint returns immediately with a conversation ID while encoding continues in the background. New submissions are queued and processed sequentially.

POST /v1/{space_id}/formation
Authorization: Bearer <token>
Content-Type: application/json

Request:

{
  "messages": [
    {
      "role": "user",
      "content": "I prefer dark mode for all my apps. My timezone is UTC+8.",
      "name": "Alice"
    },
    {
      "role": "assistant",
      "content": "Got it! I've noted your preference for dark mode and UTC+8 timezone."
    }
  ],
  "context": {
    "counterparty": "alice_principal_id",
    "agent": "customer_bot_001",
    "source": "source_123",
    "topic": "settings"
  },
  "timestamp": "2026-03-09T10:30:00Z"
}

Response:

{
  "result": { "conversation": 1, ... }
}

Fields:

Tips for best results:

• Include the context field whenever possible — it helps the encoder associate knowledge correctly
• Send complete conversation segments, not individual messages
• Include timestamps to enable proper temporal reasoning
• The name field in messages helps distinguish between multiple users in the same conversation

Recall — Query Memory

Ask a natural language question and receive a synthesized answer drawn from the knowledge graph and conversation history.

POST /v1/{space_id}/recall
Authorization: Bearer <token>
Content-Type: application/json

Request:

{
  "query": "What are Alice's preferences?",
  "context": {
    "counterparty": "alice_principal_id",
    "topic": "settings"
  }
}

Response:

{
  "result": {
    "content": "Alice prefers dark mode for all applications and operates in the UTC+8 timezone.",
    ...
  }
}

Note: result.content is the primary contract. Additional fields may vary by model/runtime.

Query examples:

Space Status

Get statistics and health information for a memory space.

GET /v1/{space_id}/info
Authorization: Bearer <token>

Response:

{
  "result": {
    "space_id": "my_space_001",
    "owner": "principal_id",
    "db_stats": {
      "total_items": 150,
      "total_bytes": 524288
    },
    "concepts": 85,
    "propositions": 120,
    "conversations": 12,
    ...
  }
}

Wiki — Versioned Reference Documents with Citations

The wiki is the space's reference memory: policies, manuals, SOPs, API docs and FAQs stored as immutable Markdown commits (git-like), retrieved by BM25 keyword search, and quoted through verifiable wiki:// citations. Search is deterministic and LLM-free; compose answers yourself and cite the URIs.

Commit (create or update):

POST /v1/{space_id}/wiki/docs
Authorization: Bearer <token with write scope>

{
  "title": "Deployment Guide",
  "content": "# Deployment Guide\n\n## Rollback\n\nUse the previous snapshot...",
  "namespace": "engineering",
  "tags": ["sop"],
  "message": "initial import"
}

• Create: omit doc_id. Update: pass doc_id and parent_version (the current_version you read). A stale parent_version returns 409 with the current version in error.data — re-read, merge, retry.
• Committing identical content is a no-op ("idempotent": true); safe to retry and re-import.
• Content is whole-document Markdown (not a diff), at most 1 MiB after normalization (413 beyond).

Search with citations:

POST /v1/{space_id}/wiki/search

{ "query": "rollback checksum", "namespaces": ["engineering"], "top_k": 8, "mode": "chunks", "expand": 1 }

Each hit carries the matching text and a citation: { "uri": "wiki://{space}/{doc_id}@{version_id}#{start}-{end}", "checksum": "sha3-256:...", "anchor": "...", "quote": "..." }. mode: "docs" returns one best hit per document. expand (0-2, default 0) widens each hit with adjacent passages; overlapping expansions merge and the citation range widens while staying verifiable. BM25 favors exact terms (product names, error codes); reformulate keywords rather than sending full sentences.

Read progressively:

GET /v1/{space_id}/wiki/docs/{doc_id}                      → metadata + table of contents
GET /v1/{space_id}/wiki/docs/{doc_id}/content?anchor=...   → one section
GET /v1/{space_id}/wiki/docs/{doc_id}/content?start=&end=  → byte range
GET /v1/{space_id}/wiki/docs/{doc_id}/content              → full text (bounded)
GET /v1/{space_id}/wiki/docs/{doc_id}/content?version=...  → historical version

Prefer TOC → section over full reads for long documents.

Manage and audit:

GET  /v1/{space_id}/wiki/docs?namespace=&tag=&status=&cursor=&limit=
GET  /v1/{space_id}/wiki/docs/{doc_id}/versions
POST /v1/{space_id}/wiki/docs/{doc_id}/archive     (hidden from search, still readable)
POST /v1/{space_id}/wiki/docs/{doc_id}/restore
POST /v1/{space_id}/wiki/verify                    {"uri": "wiki://...", "checksum": "sha3-256:..."}
GET  /v1/{space_id}/wiki/events?kind=&doc_id=

verify answers valid, superseded (a newer version exists — it names it), or invalid. Versions are immutable, so citations never rot.

Access control (ACL labels):

Documents may carry an acl_label (set via commit, or inherited from a per-namespace default configured with update_space {"wiki_acl_defaults": {"hr": "hr-internal"}}). Space tokens may carry labels: a token with labels sees only unlabeled documents plus its granted labels — enforced as a filter clause inside the same database query as retrieval, so over-broad results are structurally impossible. Tokens without labels and CWT holders are unrestricted; anonymous readers of public spaces see unlabeled content only. Denials surface as 404 (existence does not leak); the audit log, agentic recall, and the conversations endpoints (which persist full recall runner history) all require an unrestricted token and answer 403 to a labeled one. Note: OKF bundles do not carry ACL labels (the exchange format cannot express enterprise ACLs) — imported documents inherit namespace defaults.

POST /v1/{space_id}/management/add_space_token
{"scope": "read", "name": "analyst", "labels": ["engineering"]}

Read auditing and housekeeping:

update_space {"wiki_audit_reads": true} events every external search/read (WikiQueried / WikiRead, with the real actor); agent reads stay covered by recall conversation logs. Housekeeping runs automatically after maintenance cycles and on startup: the audit log is pruned to its retention cap (the prune itself is evented), and a stale-document report is refreshed. GET /v1/{space_id}/info exposes wiki_docs / wiki_chunks / wiki_versions / wiki_queries / wiki_digested / wiki_stale_docs.

Graph bridge (WikiDigest, opt-in):

PATCH /v1/{space_id}/management/update_space   {"wiki_digest": true}
POST  /v1/{space_id}/wiki/digest

When enabled, committed wiki versions are distilled into the Cognitive Nexus: an LLM proposes subject–predicate–object facts per section, and the runtime writes them as KIP propositions whose metadata always carries source: "wiki", a wiki:// citation with checksum, and the extractor fingerprint — provenance is attached by construction, not by prompt discipline. Re-committing a document supersedes facts the new version no longer asserts (metadata.status: "superseded", metadata.superseded_by → the new version). Recall answers can therefore explain why a graph fact is believed and quote the exact source passage. The digest also runs automatically after maintenance cycles and on space startup, and each run re-verifies a sample of recorded citations. Disabled by default because it writes to the graph.

OKF interchange (requires full-scope token):

POST /v1/{space_id}/wiki/import    {"entries": [{"path": "guides/setup.md", "content": "---\ntype: Guide\n---\n\n# ..."}], "namespace": "kb"}
GET  /v1/{space_id}/wiki/export?namespace=kb

Bundles follow the OKF v0.1 convention (Markdown + YAML frontmatter; concept paths become hierarchical slugs). Unknown frontmatter keys, ordering and comments survive round-trips verbatim; re-importing an unchanged bundle is a no-op (checksum-idempotent). Export adds x_anda_doc_id / x_anda_version_id / x_anda_checksum provenance keys plus a root index.md and manifest.json, so a wiki snapshot can be reviewed with git diff and replayed into an empty space. Reserved files (index.md, log.md, non-Markdown) are skipped on import.

Integration Pattern

A typical integration workflow for a business agent (replace your-brain-host with your deployment address, e.g. localhost:8042):

1. Remember: Send conversations for memory encoding

After each meaningful conversation with a user, send the messages to Formation:

curl -sX POST https://your-brain-host/v1/my_space_001/formation \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "I work at Acme Corp as a senior engineer."},
      {"role": "assistant", "content": "Nice to meet you! Noted that you are a senior engineer at Acme Corp."}
    ],
    "context": {"counterparty": "user_123", "agent": "onboarding_bot"},
    "timestamp": "2026-03-09T10:30:00Z"
  }'

2. Recall: Query memory before responding

Before generating a response, check if relevant memory exists:

curl -sX POST https://your-brain-host/v1/my_space_001/recall \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Where does this user work and what is their role?",
    "context": {"counterparty": "user_123"}
  }'

MCP Integration

If your agent has an MCP client, connect to the HTTP service's Streamable HTTP MCP endpoint:

https://your-brain-host/mcp/my_space_001

Use the same spaceId and spaceToken you would use for REST. Pass the token as Authorization: Bearer <token>. This is the preferred setup for company or team deployments where each employee's agent is assigned a dedicated Brain space.

For local desktop or development clients, run Anda Brain as a stdio MCP server and register it with that client:

MCP_AUTH_TOKEN="$SPACE_TOKEN" \
  anda_brain mcp --space-id my_space_001 local --db ./data

Core MCP tools:

If ED25519_PUBKEYS is empty, local MCP development can omit tokens. Add --mcp-auto-create-space for stdio development or MCP_HTTP_AUTO_CREATE_SPACE=true for remote development when the target space does not exist yet; remote auto-create requires ED25519_PUBKEYS plus a CWT with write scope for the target space before creating the missing space. Set MCP_HTTP_ALLOWED_HOSTS when remote MCP is exposed behind a company domain or reverse proxy.

OpenClaw Integration

The anda-brain plugin integrates Anda Brain into OpenClaw agents, providing automatic memory encoding and a recall_memory tool — no manual API calls needed.

Prerequisites: Deploy Anda Brain and Create a Space

Before installing the plugin, you need a running Anda Brain deployment plus a spaceId and spaceToken:

1. Deploy Anda Brain — see the Quick Start guide (binary or Docker, a few minutes).
2. Create a brain space via POST /admin/create_space — the space_id you choose is your spaceId.
3. Create an API key via POST /v1/{space_id}/management/add_space_token — the returned token is your spaceToken.

Install

1. Install the plugin package:

openclaw plugins install anda-brain

2. Update anda-brain configuration in openclaw.json with the spaceId and spaceToken obtained from the console:

{
  "plugins": {
    "entries": {
      "anda-brain": {
        "enabled": true,
        "config": {
          "spaceId": "my_space_001",
          "spaceToken": "STxxxxx",
          "baseUrl": "http://localhost:8042" // your Anda Brain deployment URL
        }
      }
    }
  }
}

3. Restart OpenClaw Gateway.

openclaw gateway restart

Required fields:

• spaceId: your Brain space ID (created via POST /admin/create_space)
• spaceToken: your space API Key (created via POST /v1/{space_id}/management/add_space_token)
• baseUrl: your Anda Brain deployment URL (e.g. http://localhost:8042) — always set this; the legacy default https://brain.anda.ai has been discontinued

What It Does

Configuration Options

recall_memory Tool Parameters

The plugin registers a recall_memory tool that the LLM can invoke:

Anda Bot

Anda Bot is a complete, open-source AI agent built on Anda Brain, using Brain as its long-term memory and cognitive backbone. Use it directly, or as a reference implementation for integrating Anda Brain into your own agent.

Troubleshooting

Configuration Reference

The service is configured via CLI arguments and environment variables:

Storage backends: