Hybrid search, verbatim recall, knowledge graph, and passive capture for Claude Code — local-first, and running in private beta. Works on Mac, Linux, and cloud. No hardware binding.
Every team using Claude, Cursor, or Copilot has hit these four walls. team-intel answers each one with shipped code.
“The model silently loses critical decisions made earlier in the conversation, then keeps generating with full confidence.”
Answer: hybrid search + verbatim recall. Every message is indexed in parallel with distilled extraction. When the top RRF score drops below the calibrated threshold, the RAG pipeline transparently falls back to raw message search. You search what was actually said, not just what was distilled.
“Even when you're wrong. Even when you contradicted yourself yesterday. The model has no memory of either.”
Answer: heuristic + gotcha artifacts retrieved by the same RAG pipeline. Saved insights push back with the user's prior corrections, not generic advice. A built-in reconciliation skill audits Claude memories against the knowledge base for drift.
“Frontier models cannot form long-term memories. Every new session, your team's hard-won context resets to zero.”
Answer: a passive capture hook records every tool use in under 10ms. A stop hook runs the extraction pipeline async, with a circuit breaker after three LLM failures. Knowledge survives across sessions, repositories, and machines.
“The agent will happily merge insecure code, ignore your style guide, and forget last week's architectural decisions.”
Answer: shipped skills. /team-intel-backfill extracts institutional knowledge from git history. /refresh-claude-md audits CLAUDE.md against the actual code. /goodnight is an end-of-day sweep for uncommitted work and stale knowledge. Enforcement lives in skills and hooks, not LLM “please follow these rules” prose.
Knowledge capture that stays out of your way.
Write code, debug, review PRs. Team Intel's capture hook runs in the background — less than 10ms overhead per tool call.
< 10ms overheadWhen your session ends, an LLM extracts structured knowledge: decisions you made, heuristics you discovered, playbooks you followed.
Hybrid search (semantic + keyword + Reciprocal Rank Fusion) surfaces the right knowledge when you need it. Fresh results, not stale docs.
A complete knowledge management toolkit that integrates directly into your coding workflow.
Semantic embeddings + BM25 keyword search fused with Reciprocal Rank Fusion. Get the best of both worlds — meaning and precision.
A SQLite graph with temporal edges maps relationships between decisions, people, and projects. Discover connections you didn't know existed.
Every artifact carries a freshness score — 90-day decay, nudged up when it’s used or validated — returned with each search result so agents can weigh old advice accordingly.
Works with Ollama (free, local), Claude API, or deferred extraction. Your choice of model, your data stays where you want it.
This is a real knowledge graph generated from development sessions. Drag nodes, explore connections.
Knowledge should outlive Slack threads and quit notices.
Know who knows what. Identify expertise gaps before they become crises. Every decision is linked to its author and context.
New team members can search what the team already decided before interrupting a senior.
Decisions outlive the people who made them. “Why did we choose Postgres?” is one search away.
No roadmap items. Everything below is in the codebase today.
| Dimension | team-intel |
|---|---|
| Install today | Private beta — Mac, Linux, cloud, no hardware binding |
| Source availability | Elastic License 2.0. Source shared with beta users; read, modify, and use it for yourself or your team |
| Test coverage | 1,200+ tests across core, MCP, and platform |
| Persistence | SQLite + ChromaDB (individual) or Postgres with tenant isolation (team) |
| Knowledge graph | SQLite with temporal edges, entity registry, person aliases, merge primitives |
| Hybrid search | Semantic (Qwen3-Embedding-8B) + BM25 fused with Reciprocal Rank Fusion (k=60) |
| Verbatim recall | Every session message indexed in parallel with extraction. Calibrated RRF fallback when distilled artifacts miss |
| LLM flexibility | Ollama, Claude API, or any OpenAI-compatible endpoint (vLLM, LiteLLM), plus deferred extraction |
| Pricing | Free for individuals during the beta |
Leave your email and I’ll reach out with access.
No spam. Beta invitations and launch updates only.