vectorize-io / hindsight
Hindsight: Agent Memory That Learns
github.com/vectorize-io/hindsightIndexed at ccfe85bUp to date with upstream
About this codebase
Hindsight is an agent memory system end-to-end platform: it consumes agent events (observations and retainable facts) and repository context, runs a retain/consolidate pipeline to build a bank of stored evidence, and outputs queryable memory plus per-bank Model Context Protocol endpoints that clients and agents can use to retrieve and ground knowledge over time. It is built for developers who want “learn, not just remember”—so RAG, semantic search, and onboarding flows can cite what the system believes and why, rather than only retrieving raw conversation snippets.
Written by gpt-5.4-nano from the index
4,174 files and 41,702 symbols in 19 modules, led by Python (1,904 files) and TypeScript (688).
Start reading
- Knowledge graph4,174 files and how they depend on each other.
- Admin Control EndpointsDocumentationDocumentation for hindsight-api-slim/hindsight_api/admin, written by gpt-5.4-nano.
- Do not maintain memory_units/mental_models vector/text indexes for custom store-owned memoriesDecisionProposed decision, from a pull request.
Code health
This codebase scores 6.0 out of 10 for code health, which we rate fair. It also scores maintainability 6.7 and static performance 9.7 out of 10. The three are scored separately and never blended into one number. Risk is concentrated, as it usually is: 500 of 4,174 files are git hotspots, and they average 3.9 — which is where the fixes pay off most.
On the leaderboards:#3 of 4 Agent memory layers#37 of 44 Python repos
Do next
1 thing worth doing in the week to Sep 26, the last indexed commit, 1 of them now.
Now
- Critical
Clean up what this week's commits left: 21 serious findings in 14 files
9 commits in the last 7 days added or worsened them, and they are still open. Start with
hindsight-control-plane/src/components/prompt-preview-dialog.tsx, which has 1 critical.- serious findings
- 21
- critical
- 1
- files
- 14
Large effort
Improve what Repowise can see
Add a test coverage report
Without one, Repowise cannot tell tested code from untested code, so every test-related action says “unknown”.
- line coverage
- Unknown, no report
Small effort
Use vectorize-io/hindsight in your agent
Point Claude Code, Cursor, Codex or VS Code at this repository. Your agent gets the index this page is built from: cited answers, callers, history and health for any file.
https://api.repowise.dev/mcp/vectorize-io/hindsightOne command, run anywhere. Adds the server to your local Claude Code config.
Docsclaude mcp add --transport http repowise https://api.repowise.dev/mcp/vectorize-io/hindsight \ --header "Authorization: Bearer rw_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
This repo is public, so anyone with a Repowise account and an API key can connect.
Try asking
- get_overview()
- get_answer("how is hindsight put together?")
- get_why("why is hindsight structured this way?")
Prefer local? pip install repowise && repowise init indexes your own checkout, no account needed.
Go deeper
- OverviewEverything the index found, what changed this week, commits and decisions.
- DocumentationModule-by-module documentation, 4,908 pages, 74 written by a model.
- ArchitectureDependency graph, layers and third-party dependencies.
- Code healthEvery file's score, the findings behind it, tests and refactoring plans.
- CommitsEach commit's effect on code health, with agent provenance.
- ContributorsWho owns what, and where one person holds most of the knowledge.
- DecisionsArchitectural decisions mined from commits and pull requests.
- Ask this codebaseSign inGrounded answers over the indexed code.
Maintain this repo? Claim the page
Sign in with GitHub and we check that you are an admin or maintainer of vectorize-io/hindsight. Then add a badge to the README that links readers here.
Use this repo from Claude or ChatGPT
Add this URL as a custom connector and sign in with repowise when asked. No API key to copy.
https://api.repowise.dev/mcp/vectorize-io/hindsightFor maintainers
Add a badge to your README
Every badge is public, cached, and updates on its own after every index. Nothing to install.
Repowise badge
Links to this page, with no score on it. Adding it also re-indexes the repo every week.
[](https://repowise.dev/repo/vectorize-io/hindsight?src=badge_wiki)Code health badge
The repo's health score out of 10, from the latest index.
[](https://repowise.dev/repo/vectorize-io/hindsight?src=badge_health)Code health card
A larger card with the score, its change since the last index, and when it was measured. The HTML version follows the reader's light or dark theme.
[](https://repowise.dev/repo/vectorize-io/hindsight?src=badge_card)How vectorize-io/hindsight works
Hindsight: Agent Memory That Learns This page is a map of the vectorize-io/hindsight repository, written primarily in Python, rebuilt from the source each time it is indexed. Repowise parses every symbol, computes a dependency graph, scores per-file code health from complexity, duplication, test coverage and churn, mines git history for hotspots and ownership, and lifts the architectural decisions into documentation you can read here or query through MCP.
The codebase has 4,174 files and 41,702 symbols in 19 modules, led by Python, TypeScript and Go. Code health is 6.0 out of 10, rated fair.
Use the links above to open each view, or connect this repository to your agent for grounded answers inside Claude, Cursor, Codex or VS Code.