Anvaya Mind · THE EXPERIENCE LAYER
Experience ThatCompounds AcrossEvery Session.
Most memory layers give an agent a place to store facts. Mind is built around a different idea: the lifecycle a junior engineer actually goes through on the way to senior. It captures what was decided, what broke, what pattern held, the dos and don’ts and the domain knowledge the repo never states. It runs locally, stores everything under .anvaya/mind/, and hands back a packed context sized to whatever budget you give it. It also exposes that experience to Claude Code, Cursor, or any MCP client through an MCP server, so the graph serves the agent you already use.
The Shape Of It
Not A File. A Graph.
Six node types, each with its own decay constant. A decision about your auth model should outlive the bug fix you shipped last Tuesday, so it does, by roughly fifteen times.
Six Stages Between Question And Context.
Every query runs the same pipeline. The expensive part isn’t finding related nodes, it’s deciding which ones still deserve your token budget.
Where The 12–25x Estimate Comes From.
Per-node compression is a design estimate from Anvaya’s memory design work, each node type stores a fraction of the raw tokens that produced it, not a benchmark result. The measured end-to-end lever today is −11.0% median uncached tokens (micro-fold, n=5, verification 19/19); Phase 6 measures the compounding curve itself.
Experience That Compounds.
A memory layer stores what you tell it. Anvaya Mind captures what a project teaches you: the debugging session, the decision, the pattern, the dos and don’ts, the domain history the repo never writes down. Every session feeds the graph, six node types each decaying on its own curve, with drift checked against git, so the agent compounds toward seniority instead of restarting cold. The compounding claim (session N+1 costing less and knowing more) is the thesis Phase 6 of the benchmark program exists to measure.
Mind is empty. Every query requires full file reads, no different from any other AI tool.
First bug fixes and decisions captured as typed nodes. Retrieval starts replacing repeat reads.
Decisions, patterns and bug fixes accumulate with per-type decay. Causal-thread materialization is implemented but default-off, so this stage is a design target until it is promoted.
The thesis: a solo developer's graph holds the decisions and failed approaches a team would remember. No longitudinal study exists yet: Phase 6 of the benchmark program measures whether session N+1 costs less and knows more than session N.
Everything Local, Everything Inspectable.
One directory inside your project, gitignored by default. SQLite in WAL mode for nodes, links, sessions and full-text search; a separate file for the Int8-quantized HNSW index. Nine background workers and an optional MCP listener keep it consolidated.
Run Agents That Fit On Your Laptop.
25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.
Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.