FAQ · STRAIGHT ANSWERS
If you’re evaluating this for real, these matter more than the feature list.
The Questions
24 concurrent by default, with a 256-agent hard cap and a 20-per-second spawn rate. It is light enough that the machine, not the software, is usually the limit: a 32-agent soak cost 103 MB total agent RSS, and 25 sessions ran in parallel on a Core 2 Duo with 4 GB of RAM, no crashes, no faults. On a modern laptop, roughly a hundred agents fit inside about 1.5 GB of harness memory.
No. Anvaya Mind runs as a local daemon and stores everything under .anvaya/mind/ inside your project, SQLite, an HNSW vector index, and an append-only temporal log. On the Ollama path, the entire loop runs on your machine and nothing is transmitted. Anthropic and OpenAI-compatible endpoints are BYOK cloud paths you opt into.
Yes, over MCP. Anvaya ships an MCP memory server, so any MCP-capable client can query the same memory. A five-line .mcp.json in your project root exposes fourteen tools: mind_query, mind_search, mind_record_change, mind_get_node, mind_feedback, mind_stats, mind_start_session, mind_end_session, mind_get_threads, mind_proactive, mind_delta, mind_intent, mind_retro_promote, mind_federate.
Three transports: Ollama (local, no account or key), generic OpenAI-compatible endpoints (OpenAI, Siemens preset, vLLM, LiteLLM, LM Studio, OpenRouter, custom gateways), and Anthropic Messages. The Ollama path is first-class: embeddings, summarization, and generation all run locally with a pulled model like devstral.
A CLAUDE.md is flat text you maintain by hand, it grows, rots, and keeps asserting the old thing until someone edits it. Mind is a weighted, linked, decaying graph: nodes have types and half-lives, stale ones are detected against git drift, and only the relevant few are packed into each query's token budget.
No, and that's deliberate. The graph is project-local and tied to one machine. There is no cloud sync and no multi-machine replication, the files live in your repo's working tree, gitignored by default.
Mind is a Unix-socket daemon, so the supported surface today is Linux and macOS. The CLI builds from source with Rust 1.75+ and runs as a TUI or headless in CI.
Public beta launches in Q4 2026. Anvaya is pre-1.0 and unreleased, both halves build from source, and the changelog is a real build log. Pricing isn't set, but the core will stay local-first regardless of what the SaaS layer ends up looking like. Early access is open via the waitlist.
Claude Code uses CLAUDE.md, a flat markdown file you write and maintain by hand. It can grow large, it asserts stale information until someone edits it, and it doesn't learn which parts are actually useful. Anvaya Mind is a weighted, linked, decaying graph: nodes have types and half-lives, stale content is detected against git drift, and only the relevant few are packed into each query under a token budget. The key difference: Claude Code's memory is static. Anvaya's improves with use.
The memory graph is model-agnostic. It stores typed nodes, links, and utility scores, not model-specific embeddings. Switching from Ollama to Anthropic (or vice versa) doesn't invalidate the graph. The daemon continues using whatever embedding provider you configure.
Yes. The graph is files: SQLite in WAL mode, a flat HNSW index, an append-only temporal log, and plain-text config, all under .anvaya/mind/ in your project. You can open the SQLite database directly, query it, export to CSV, or back it up. There's no lock-in; the data format is documented and inspectable.
Anvaya is independent and pre-1.0. It was selected for the Sarvam AI Startup Program in 2026, which grants API credits, mentorship and priority support, not investment, not equity, and no board seat. It does not make Anvaya cloud-dependent: Anvaya Mind runs as a local daemon under .anvaya/mind/ in your project, and the Ollama path needs no account, key or cloud.
Glossary
Technical terms explained in plain language. Hover over abbreviations on other pages for quick definitions.
A statistical model that updates beliefs based on observed outcomes. In Anvaya, it scores whether each memory node actually helped the AI, nodes that are used get higher scores, unused ones decay.
Hierarchical Navigable Small World, a fast algorithm for approximate nearest-neighbor search. Anvaya uses it to find semantically similar memory nodes quickly, without scanning the entire graph.
Content-anchored checking against git history. When code changes, Anvaya detects which memory nodes are now stale and flags them for decay or removal.
The ratio of raw context size to the packed context Anvaya serves. Instead of feeding the AI 50K+ tokens of raw files, Anvaya selects the 2-4K most relevant tokens, a 12-25x compression (design estimate, not a benchmark run).
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.