Show HN: CtxVault – receipts for AI context, not another memory store
CtxVault provides receipts for AI context, allowing teams to inspect and trace the evidence that shaped an AI's output. A case study demonstrates how from 2056 candidate refs only 478 were selected, improving auditability.
Then what?
Before
Agents pull from memory, docs, tickets, and local files.
Reviewers cannot tell what actually shaped the next answer.
With CtxVault
The AI-facing packet carries selected, caveated, blocked, and omitted evidence.
The team can inspect the receipt before trusting the output.
After
Unsafe or unsupported context has a visible reason.
Bad projections have a rollback target instead of a blame trail.
The control point
Source refsPublic or local evidence starts as candidate context.
Review decisionRefs are selected, caveated, blocked, omitted, or marked unsupported.
Context packetReviewed material can become an AI-facing packet such as AGENTS.md or CLAUDE.md.
ReceiptHashes, side effects, omissions, and rollback path stay inspectable.
mem0 read-only governance case study
The case study makes the pain visible: from 2056 candidate refs, only 478 were selected, 263 were caveated, 98 were blocked, and 1217 were omitted from the projected context. This is about governing influence, not storing more memory.
2056candidate refs
839surfaced refs
478selected refs
263caveated refs
98blocked refs
Selected
Public docs and protocol refs.
Evidence precision: 0.570.
Caveated
Public but scope-sensitive generated or workflow refs.
Credential-shaped examples stay bounded.
Blocked / omitted
Media assets and credential-shaped markers are not promoted.
Omitted samples are visible up to the receipt cap.
This public case study uses a sanitized extract over a pinned public OSS receipt. It does not run mem0 code, call a provider/model, execute an adapter, write target files, judge project quality, or imply maintainer endorsement.
Read the evidence
Deep dive · Public JSON extract · GitHub repository