daidocs

daidocs is an open-source AI memory system built by Kerneta (a trading name of Siro Robotics Ltd, registered in England and Wales) that stores LLM conversation history as plain-text .dai files on your own machine. The project launched its V4.4n32 release on 18 September 2026 under the Apache 2.0 licence, with the complete engine, MCP server, and benchmark artifacts published in the same repository.

What It Is

.dai is a file format — not a memory service — designed to make AI assistant memory portable, inspectable, and vendor-independent. Each conversation becomes a single plain-text file with three zones: a YAML header (identity and metadata), a fenced JSON block (machine-parseable facts, events, and entities), and the cleaned original content. A small derived index sits beside the files for fast retrieval. Because the store is a folder of text files, it works with grep, git, any editor, and any model that speaks MCP — Claude Desktop, Claude Code, Cursor, Windsurf, Codex, Cline, Continue, Zed, the OpenAI Agents SDK, and local models.

How the Memory Architecture Works

The system separates two model roles: an observer that converts a conversation into a .dai file once, and an actor that answers questions from the store. Conversion is a single model call per conversation; recall is index lookup and file reads with no model call at all. The retrieval layer reads a question-specific slice of the store in three zooms — manifest first, then a file's Understanding block, then specific content segments only if needed — rather than loading the whole history. According to Kerneta's published benchmark run on LongMemEval-S (500 questions, GPT-4o answering), the engine reads a mean of 10,065 tokens per question against histories averaging 103,601 tokens, approximately 10.3× fewer input tokens than pasting the full history.

Benchmark Position and Accuracy Claims

Kerneta publishes the following figures, all attributed to LongMemEval-S with GPT-4o answering and the benchmark authors' own evaluate_qa.py scorer:

• 83.00% (415/500) with GPT-4o as the answering model — second among memory systems whose configuration is reproducible by someone outside the vendor
• 92.00% (460/500) with Claude Fable 5 as the answering model
• 60.60% (303/500) for the same GPT-4o model with no memory system (full history pasted), the benchmark authors' own baseline
• The system ahead on the same protocol is Mastra Observational Memory at 84.80%, which Kerneta states reads approximately 30,000 tokens per question versus their 10,065

The replication protocol is published in docs/REPLICATION.md in the repository, and per-question judge verdicts are included in benchmark/.

Deployment Model and Surfaces

The complete local product runs free and self-hosted. npx daidocs setup detects and configures Claude Desktop, Claude Code, Cursor, Windsurf, Codex, Cline, Continue, and Zed in one command, installs session hooks, and backs up every file it touches. On a Claude subscription, no API key is required — the assistant in the session writes each memory itself. A Python reader (pip install daidocs) provides pure-Python access to the store without Node. Hosted surfaces — the claude.ai web connector, mobile (iOS/Android), team vaults with SSO, and org-wide provisioning — are described as in waitlist and not yet available.

Update: V4.4n32 Launch Release

The repository went public on 13 September 2026, with the V4.4n32 launch release published on 18 September 2026. This is the initial public release of the engine, format specification, MCP server, benchmark adapter, and all run artifacts. The GitHub README notes the repository was published only after a full key rotation and history scrub. The format version is 4.4, and the engine designation is v4.4n (the "n" variant that produced the published 83.00% figure). Earlier engine versions appear only as aggregate totals in the version ladder on the results page.

Open-Source Model and Funding

The format specification, the v4.4n engine, and the MCP server are all Apache 2.0, including a patent grant. Kerneta states the local product will never be crippled to upsell a hosted tier — cloud plans sell hosting, sync, and team features, not gated engine capabilities. The project is self-funded with no external investment at the time of launch, according to the pricing page. Supporter badges (one-off payments that convert to indexing credit) are the primary early funding mechanism alongside pay-as-you-go hosted conversion.