Living Brain

Living Brain is a hosted agent memory service from ChatAndBuild. It stores information outside a large language model, organizes it into connected knowledge pages, and makes selected material available to assistants and custom agents across sessions. Its developer platform provides an API, SDKs, a command-line interface and Model Context Protocol (MCP) integrations. The API documentation identifies the service as being in early access.^[1][2]

Overview

Living Brain separates stored knowledge from the model that answers a question. A user can accumulate notes, documents and conversations in a brain, then let a connected assistant retrieve that material. The documentation describes a brain as a knowledge store belonging to an end-user subject, with each capture, compiled page and connection assigned to one brain.^[3]

The product's compiled pages and connections form a knowledge graph. Its memory is supplied to a model at inference time. Retaining a new fact in these pages does not itself change that model's weights.^[1][3]

Capturing and maintaining knowledge

Capture processing is asynchronous. An initial request returns an ingest-source record; background processing extracts information, chooses an existing page or creates a page, records a version with source provenance, and refreshes retrieval data and connections. A source can report a failure even after a page has been saved if a later processing step fails.^[3]

The capture interface accepts batches, with each item processed independently. A separate chat-turn operation uses a classifier to decide what should be retained from a conversation. Failed sources can be retried. Deleting an ingest source does not automatically delete pages already derived from it.^[4]

Pages can be active, stale, contested, deprecated or archived. The documentation describes marking conflicting material as contested and preserving the competing claims for resolution. Brain-profile changes affect subsequent captures rather than retroactively recompiling existing pages.^[3]

Memory policies

Living Brain documents four internal policies for deciding how to maintain and reuse information.^[5]

The published policy descriptions include eligibility and permission checks before ranking. Their numerical weights and thresholds are not published, and the API does not return the policy scores. These scores express maintenance or retrieval decisions rather than probabilities that information is true. The documentation also distinguishes resurfacing relevant material from granting an agent permission to act.^[5]

Retrieval and prompt context

Living Brain combines semantic search with keyword matching. Semantic retrieval can find differently worded material, while keywords help preserve matches for names and identifiers. Search results indicate whether a match was semantic, keyword-based or both.^[6]

The documented defaults return five results and apply a semantic-similarity floor of 0.35; keyword matches can remain below that floor. Search embeddings cover the title, summary and beginning of a page's body. A result's similarity field is a vector score, not a probability that the page answers the question, and it is separate from the combined ranking.^[6]

For general session context, a prompt-index operation returns a compact selection of important pages, with pinned pages placed first. An agent can use this index at the beginning of a conversation and search for narrower evidence when needed. The graph API returns nodes and edges for an application to render; the sign-in MCP connector also provides an interactive graph inside compatible chat clients.^[7]

This is a form of external memory retrieval related to retrieval-augmented generation. The maintained pages add a compilation stage before a later query, rather than requiring the answering model to assemble all related raw passages for each question.^[3][7]

Developer interfaces

The API-key MCP documentation includes integrations for Claude Code, OpenAI Codex, Cursor and Visual Studio Code. It maps API operations to tools for captures, brains, workspace content, integrations and webhooks. The sign-in connector is a separate route for compatible clients, including ChatGPT and Claude; the tools exposed depend on the OAuth scopes granted.^[9][13]

API requests use an organization-scoped key and an x-subject-id identifying the end user. The key inherits the issuing member's role. The authentication guide instructs developers to keep organization keys on the server and use a backend proxy for browser access.^[8]

Imports and synchronization

The documentation provides a Markdown-vault importer and workflows for bringing in Claude-project material or a manually reviewed export of an assistant's remembered information. These workflows submit captures for compilation; connecting the service does not automatically copy every conversation or every memory from all other assistants.^[14]

The importer uses source identifiers to avoid resubmitting unchanged notes. Deleting or renaming an original vault file does not remove previously compiled knowledge. Its documented text workflow also leaves binary attachments behind unless they are captured separately.^[14]

Applications can follow a change feed or subscribe to webhooks for page changes, new connections, completed briefs and import progress. Daily briefs can run on a schedule or be requested directly. Acknowledging a change marks it as seen by a user; machine synchronization separately tracks its position in the feed.^[15]

Reported benchmark results

Living Brain publishes internal results on conversational-memory benchmarks. As inspected on November 5, 2026, its homepage and developer benchmark page reported different LoCoMo figures. The dated documentation run and the homepage figures should be treated as separate reports.^[1][16]

The documentation identifies the October 26 run as batch 4fd3f746, with binary judgments against reference answers. Its full-history figure is a context-size comparison, without a reported accuracy result for that baseline. It also describes the figures as coming from one run.^[16]

Separately, the homepage reports 22.6 times fewer tokens and 95% lower cost per task on LongMemEval-S against a 116K-token full-context baseline. It reports zero observed hallucinations under that evaluation's scoring criteria. Those observations do not establish that the service never produces false information.^[1]

Interpreting the evaluations

The original LoCoMo research evaluates long conversational histories through question answering, event summarization and multimodal dialogue generation. Its published question-answering results use an F1 metric. Living Brain's binary judged accuracy should therefore not be treated as directly interchangeable with the original paper's F1 scores.^[16][17]

LongMemEval is a separate research benchmark covering information extraction, reasoning across sessions, temporal reasoning, knowledge updates and abstention. Its existence and design do not independently validate Living Brain's reported implementation results.^[18]

The Living Brain benchmark documentation explicitly limits its LoCoMo evaluation to answering questions over a fixed history. It does not measure whether the service proactively resurfaces information at the right moment or whether people find its compiled pages useful. Internal rankings apply to the tested systems and configurations.^[1][16]

Continue reading

• AAgent memoryAgent memory is the set of systems that let an AI agent retain and recall information beyond a single context window, so it can carry knowledge, preferences, and experience from one interaction or session into…Linked in the introduction
• Large Language ModelA large language model (LLM) is an artificial intelligence system built on a transformer neural network with billions to trillions of parameters, trained on massive text corpora to predict the next token in a…Linked in the introduction
• Model Context ProtocolModel Context Protocol (MCP) is an open protocol for exchanging context and invoking capabilities between applications that use large language models and external programs or data sources.Linked in the introduction
• Claude CodeClaude Code is an agentic software development product made by Anthropic. It uses models from the Claude family to inspect software projects, propose and apply changes, run development tools, and report…Also in Developer Tools
• Cursor (code editor)Cursor is a proprietary AI-assisted code editor and software-development platform made by Anysphere.Also in Developer Tools
• OpenAI CodexOpenAI Codex is the brand OpenAI has used for two distinct generations of code-focused artificial intelligence products: a 2021 large language model that turned natural-language prompts into codeAlso in Developer Tools

More to read

• n8n
• Muse Code
• Pi (agent framework)
• Mistral Vibe
• Gumloop
• Lindy

Topics:AI AgentsAI Tools & ProductsDeveloper Tools

Related topics

• Agent memory
• Knowledge graph
• Model Context Protocol
• Retrieval-augmented generation

References

1. ^1 ^2 ^3 ^4 ^5Living Brain. Official product website. Product description, FAQ and internal benchmark statements; accessed November 5, 2026.
2. ^Living Brain. Developer documentation overview. Early-access status and integration overview; accessed November 5, 2026.
3. ^1 ^2 ^3 ^4 ^5Living Brain. How a brain works. Capture pipeline, page lifecycle and knowledge connections; accessed November 5, 2026.
4. ^Living Brain. Capturing knowledge. Capture operations and ingestion behavior; accessed November 5, 2026.
5. ^1 ^2Living Brain. Design policies. Memory maintenance and resurfacing policies; accessed November 5, 2026.
6. ^1 ^2Living Brain. How search ranks pages. Search ranking and score interpretation; accessed November 5, 2026.
7. ^1 ^2Living Brain. Search and prompt context. Retrieval, prompt index and graph interfaces; accessed November 5, 2026.
8. ^Living Brain. Authentication. Organization keys, subject identity and roles; accessed November 5, 2026.
9. ^1 ^2Living Brain. MCP with API key. Hosted MCP transport and tools; accessed November 5, 2026.
10. ^Living Brain. TypeScript SDK. Package and runtime requirements; accessed November 5, 2026.
11. ^Living Brain. Python SDK. Package, clients and runtime requirements; accessed November 5, 2026.
12. ^Living Brain. CLI. Command interface and output formats; accessed November 5, 2026.
13. ^1 ^2Living Brain. MCP with sign-in. OAuth connector and scope-dependent tools; accessed November 5, 2026.
14. ^1 ^2Living Brain. Importing existing knowledge. Vault, assistant-memory and project imports; accessed November 5, 2026.
15. ^Living Brain. Change feed and webhooks. Synchronization and daily briefs; accessed November 5, 2026.
16. ^1 ^2 ^3 ^4Living Brain. Benchmarks. LoCoMo batch `4fd3f746`, October 26, 2026; accessed November 5, 2026.
17. ^Maharana, Adyasha, et al. Evaluating Very Long-Term Conversational Memory of LLM Agents. LoCoMo research project, 2024.
18. ^Wu, Di, et al. LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. arXiv:2410.10813, October 14, 2024; ICLR 2025.

Show all 18 references