Your team uses Claude Code? Now it remembers your codebase.
Persistent neural memory for AI agents. Your agent opens the right file first — 93.75% gold-file recall across 40 pre-registered queries on four public repos, at 45–257× fewer tokens than pasting those files in. Local-first, no telemetry.
- 93.75%
- Gold-file recall
- 45–257×
- vs. pasting files
- 40
- Pre-registered queries
- Auto-provisioned
- Free tier
mean, 79–100% per repo
fewer tokens
four public repos
no signup
How it works
Three steps, then it runs forever on every git commit via neuralmind init-hook .
- 01 / 03
Index
tree-sitter parses your codebase into a graph — functions, classes, imports, comments. Ten languages, zero config. TurboVec compresses vectors 4-bit.
~15 min - 02 / 03
Query
Your agent asks a question. Progressive L0–L3 disclosure pulls exactly the right amount of context — a local index lookup, not another model call.
local lookup - 03 / 03
Remember
A Hebbian synapse layer learns co-activations from how you actually use the codebase. Budget-neutral, runs in the background.
Forever
Benchmarks
neuralmind benchmark . for your own.- Gold-file recall
93.75%
40 pre-registered queries, 4 pinned OSS repos (79–100% per repo)
- Tokens vs. pasting files
45–257×
same 40 queries; cheaper than ripgrep on every repo
- Learned recall
Never worse
CI asserts synapse recall ≥ no-recall on the same warm graph, at a neutral token budget
- vs. naive truncation
Never worse
CI asserts our selection beats truncation at an equal budget. Both magnitudes vary by repo — run them for yours
- Field report, one repo
48.8×
~9,300-node private TypeScript codebase — method reproducible, not CI-gated
- Setup time
~15 min
one CLI command; post-commit hook keeps it current
Plus one labeled field report (measured with the CLI, not CI-gated): 48.8× on a real ~9,300-node TypeScript SaaS platform, through a major rebuild. Read the field report →
See it in 30 seconds
Clone the repo, run the demo, get numbers on YOUR codebase. Then email the output to hello@neuralmind.uk with your team size for a free full spend model.
Prove it in 5 minutes
30-second demo on a fresh clone
One script: isolated venv, index build, three real questions against the bundled fixture project.
$ git clone https://github.com/dfrostar/neuralmind && cd neuralmind
$ bash scripts/demo.sh
Then measure YOUR codebase
The fixture is tiny (~500 lines, ~5.5×). Real repos consistently hit 12–50× on the same pipeline — run the benchmark on your own code and read your own number. A maintainer field report on a ~9,300-node private TypeScript codebase measured 48.8×.
$ pip install neuralmind
$ cd /path/to/your-repo
$ neuralmind build .
$ neuralmind benchmark .
Verify what you installed
Check the SBOM, release integrity, and audit trail on the security page — and the full production before/after data on the effectiveness page.
What the demo prints (bundled fixture)
Q: How does authentication work in this codebase?
naive = 4,736 tok neuralmind = 829 tok reduction = 5.7×
Average reduction: 5.5× across 3 queries
Avg context size: 859 tokens (vs 4,736 naive)
Small fixture, small multiplier — by design. It runs in CI on every commit as a regression gate. The 12–50× headline comes from real repos; your own number is one neuralmind benchmark . away.
The savings are free. The tier is control.
The savings are free
- Token compression ships in the free MIT core — the savings cost nothing, and you can measure them on your own repo in ~15 minutes.
- Modeled at 30 code questions per developer per day, a 50-developer team gets back ~$310/mo on inference alone. The figure barely moves with the exact ratio — 48.8× and 65.6× differ by under 1% of the saving, because both already remove ~98% of the tokens.
- The bigger line is time: ~$1,650/mo per 50-dev team recovered from context-limit thrashing and re-prompting, at a $50/hr fully-loaded rate.
- Recall is a local index lookup, not another model call — nothing in the loop between "I need to know X" and "I know X" waits on an API. That compounds across every developer, every day.
Fewer wrong answers, faster teams
- 93.75% gold-file recall across 40 pre-registered queries on four public repos — 79–100% per repo, with every miss published rather than dropped.
- Team dashboard shows synapse memory health, ingestion status, savings, latency trends — all read-only, all local.
- Self-documenting code: DocEvolver finds undocumented methods and evolves JSDoc that actually improves retrieval.
- The engine transmits no repository content and sends no telemetry — verifiable on the wire. Works with the agents you already run: Claude Code, Cursor, Cline, any MCP agent. No rip-and-replace.
Features
Everything persistent memory should be — and nothing it shouldn't.
Local Retrieval, No Model Call
Recall is a local index lookup, not another round trip to a model. The ChromaDB-free TurboVec backend keeps 4-bit quantized vectors 8–16× smaller with fact recall 0.800 vs 0.744 for float32 — parity gated in CI.
Hebbian Synapse Layer
Associations strengthen when you use them together — like a real hippocampus. Budget-neutral, no cost until it activates.
Progressive L0–L3 Disclosure
Retrieves the exact bytes needed. Never pastes the whole repo. 45–257× fewer tokens than full-file context across 40 pre-registered queries on four public repos.
Team Dashboard
Read-only web UI: synapse memory, ingestion status, savings, latency trends, community distribution, recent queries.
Self-Documenting Code
DocEvolver finds undocumented methods, generates JSDoc variants, evolves them against retrieval fitness. Winning variants patched back into source.
Business-Context Synapse Seeding
seed_from_documents() builds deterministic, LLM-free associations between business documents (decisions, SOPs, meeting notes) and your code graph. Adjacency-matched compounds, title-reference cross-links.
MCP Server
First-class MCP integration. Works with Claude Code, Cursor, Cline, Continue, and any MCP-compatible agent.
Ten-Language Code Graph
tree-sitter indexes Python, TypeScript, Go, Rust, Java, C, C++, C#, Ruby, and PHP out of the box.
Tool Output Recovery
Caches dropped tool output from context windows. When your agent forgets, neuralmind remembers.
Free Tier — Auto-Provisioned
pip install neuralmind && neuralmind wakeup . writes the license on first run. Zero signup wall. Default tier is "free", identity auto-issued.
100% Local Engine
NeuralMind transmits no repository content — only the minimal relevant slice ever reaches your AI tool, never your whole codebase. No telemetry. One first-build model download aside, pre-seedable for air-gapped installs.
One-Command Project Init
neuralmind init auto-detects project structure, installs hooks, builds the index — all at once. Your project is ready in seconds.
Compliance Annotation Engine
Scans code and docs for annotations across CMMC 2.0, NIST SP 800-53, SOX ITGC, HIPAA, SOC 2 and ISO 27001. Each framework carries its own marker on the line — # NIST AC-1:, # CMMC AC.L2-3.1.1:. Ingest CMMC assessment guides, export control mappings, gate CI on annotation health.
Audit Export & CI/CD Check
neuralmind export --controls produces control-to-code mappings as CSV for evidence submission, or an SSP report with --format pdf. neuralmind ci-check gates builds on annotation health.
Compliance Report MCP Tool
neuralmind_compliance_report surfaces live compliance stance from any MCP-compatible agent. Ask "are we compliant on access control?" and get an answer grounded in real annotations.
See your savings before you spend a dollar
pip install neuralmind and you have the whole product. For teams evaluating at scale, we run a free AI-spend assessment: measured on your code, modeled in your numbers, no obligation.What the assessment gives you
- A measured token-reduction ratio on one of your own repos — run locally, with none of your code transmitted
- A spend model in your numbers: per-seat subscriptions, usage-based API (OpenRouter, Bedrock, Vertex), and self-hosted GPU
- A productivity model: hours lost to context-limit thrashing and re-prompting, valued at your fully-loaded rate
- An honest fit verdict — if your workload is generation-heavy or caching already covers you, we say so
The software is MIT-licensed and free. Commercial support for deployment and integration is available.
Common questions
How is this different from RAG?
RAG stores chunks and retrieves by similarity. NeuralMind learns HOW you use code — which functions you look at together, which modules correlate, which tools you run after reading what. That co-activation signal is what makes the synapse layer outperform plain RAG on real agent workloads.
Does it work with my agent?
Yes. NeuralMind exposes an MCP server, and we support Claude Code (hooks), Cursor, Cline, Continue, and any MCP-compatible agent. We also have an HTTP API and a CLI.
Does any code leave my machine?
NeuralMind itself makes no external calls — the graph, embeddings, and synapse store are all local, with zero telemetry. The only thing that leaves is the minimal context slice your AI agent (Claude Code, Cursor, etc.) sends to its own model on each query — which is exactly what NeuralMind shrinks 12-50× versus pasting whole files.
What is the CI-gated tuner?
Autopilot — shipped as v0.10.1: a population-based evolutionary search proposes configs; an independent quality harness validates against fixture queries before promotion. Architecture-complete, signal-unproven (no real production promotions yet). Promotion requires harness pass AND beating incumbent by hysteresis margin.
What does "free tier auto-provision" mean?
When you run `neuralmind wakeup .` for the first time, NeuralMind checks if a license file exists at `~/.config/neuralmind/license.json`. If not, it auto-issues a free license (tier="free", seats=1, expires="never") — no signup, no payment, no account. Your identity is created on first meaningful action.
Is NeuralMind free? What does the paid tier actually buy?
The MIT core is free forever — including all of the token compression; the 12-50× savings cost nothing. A 1-seat free license auto-issues the first time you run `neuralmind wakeup .`, no signup. NeuralMind Team ($29/user/mo, annual, 5–50 seats) adds what engineering organizations need on top: admin-controlled memory governance, a tamper-evident hash-chained audit log, seat management, and self-hosted deployment. You are paying for compliance and control, not for the compression.
What's the business case for a team?
Two lines: the measured token reduction (free, verify it on your own repo in ~15 minutes), and modeled productivity recovery — engineers stop losing hours to context-limit thrashing and re-prompting. We publish the full model with its assumptions, and the free assessment runs it in your numbers. If your workload is generation-heavy or prompt caching already covers you, we say so.
What languages does it support?
Ten out-of-the-box: Python, TypeScript, Go, Rust, Java, C, C++, C#, Ruby, and PHP. tree-sitter handles parsing.
How does team memory work?
A team's learned associations (which code goes with what) are committed to the repo as .neuralmind-team-memory.json. Every teammate's agent inherits it automatically on the next session — a new hire's agent starts already knowing "the auth handlers go with the JWT utils," instead of relearning from scratch. Zero manual steps, travels with git clone.
Why not just use Cursor / Windsurf / Aider memory?
They vendor-lock memory to their agent. NeuralMind is agent-agnostic. Memory persists across agent migrations, is inspectable, and belongs to you.
Explain "12–50× token reduction" more concretely
When your agent asks "How does auth work?", a naive approach pastes every file (~5K–30K tokens). NeuralMind retrieves L0 (symbol/structure) + L1 (docstrings/comment) + L2 (method body, selective) + L3 (only if hit-rate requires). Result: exact context needed, typically 500–800 tokens, measured in CI on every commit. The 12-50× range reflects realistic human baseline (~10K tokens); vs naive "dump all files" baseline the ratio is higher.
Stop paying for tokens you don't use.
One command to install. Seconds to verify. Your agent remembers what matters.
$pip install neuralmind- No cloud calls
- CI-verified
- 100% local