Pensieve

DrJonaC/Pensieve is ranked 10,510 among these 8 peers; regenrek/deepwiki-mcp is ranked 10,502.

Pensieve: Know How AI Remembers You **Inspect. ** A local-first workspace for structured AI memory, with a plugin-ready dashboard. Memory can shape an assistant's answers, but users need more than a hidden store: they need to inspect what was retained, understand what surfaced, and change it. Pensieve brings **observability, portability and user control** into one local interface. Pensieve manages **external structured memory**, not model weights or internal attention. The shipped adapters use Pensieve's own repository. They do not automatically

read or modify private ChatGPT, Claude, Codex or Claude Code memories. Open ** A new library starts with bundled sample memories. No API key is needed for memory management or Mock mode. For optional Live mode, create in the project root: The server uses the OpenAI Responses API. Only explicitly submitting a Live query sends the filtered query and selected memory context to the model. Restart the app after changing server configuration. This is a single-user application, **not an authenticated

public service**. Do not expose it publicly without access controls. How It Works The default store is ; can relocate runtime data, including governance reports. Retrieval uses a local lexical vector index rebuilt from current records, followed by ranking modifiers. It is not an external semantic embedding service. Hidden/deleted records are excluded from retrieval; edits replace indexed content. Providers separate memory access from host adapters and UI. Governance reports record desired state, with JSON manifests and verification receipts. The heatmap

is an illustrative local surface, **not measured model attention or causal attribution**. Model-generated explanations are interpretations, not proof. Privacy and Recovery Credential filtering is best-effort, not a complete secret or personal-data detector. It covers recognized token formats, labeled credentials, private keys and selected server-secret values. Errors retain their original language after filtering. **Raw memories, source records, exports and backups remain original, unencrypted local data. Deletion removes the active record, not every historical backup; it is not

secure erasure. Imports merge by default without silently overwriting conflicts. JSON roundtrips preserve source metadata, timestamps and governance state. Text is stored as supplied, one nonempty line per memory, without inferring personal facts. Neither export format is advertised as a native ChatGPT/Claude backup format. See migration and recovery and privacy boundaries. Verification The migration/privacy update passed 72 unit tests, 4 package-policy checks, strict TypeScript and a production build. Browser suites cover memory management, localization and credential redaction

using an isolated library and synthetic data. No real-provider call is included in those checks. Integration and Research The product contribution is a user-facing memory control surface; the engineering contribution is shared persistence and provider/host boundaries. Research questions include how users interpret retrieval, how governance affects future recall, and how to verify that user decisions were applied. Native host-memory write-back, semantic embedding providers, broader privacy classification and measured pre/post governance evaluations remain future work

. See the plugin setup reference, architecture, and project showcase. Focused contributions to providers, governance, privacy and evaluation are welcome.