DrJonaC/Pensieve

Observe how memory shapes an answer. Pensieve is an interactive system for visualizing, interpreting, and managing how Large Language Models (LLMs) “remember” a user. It bridges the gap between model-level mechanisms and user-level understanding, making AI memory observable, explainable, and partially controllable.

What it solves

AI systems often build a "working memory" of user preferences and habits over time, but this memory is typically hidden from the user. Pensieve provides a way to see, understand, and control this structured memory so users can manage what the AI remembers and why it surfaces certain information.

How it works

Pensieve acts as a host-agnostic dashboard that connects to AI memory providers. It reads structured memory records—which include metadata like priority scores, risk levels, and activation counts—and presents them in a user-facing control surface. Users can apply governance actions such as pinning, softening (lowering prominence), or hiding memories. These decisions are then compiled into a governance report (Markdown) and a machine-readable manifest (JSON), which a provider applies to the memory store to update the AI's future context.

Who it’s for

Users and developers who want transparency and agency over the long-term memory of AI agents and LLM-based systems.

Highlights

  • Governance Loop: A complete cycle of observation, user decision, and verified application of memory changes via receipts.
  • Host-Agnostic Architecture: A modular design using providers and adapters, allowing it to work with different AI hosts (e.g., Codex, Claude Code).
  • Reversible Controls: Memory management tools that allow users to pin, soften, hide, and restore records rather than just deleting them.
  • Safety Model: Includes protected display tiers for sensitive information and a verification system to ensure changes are actually applied.

Related

  • Project
    atomicstrata/atomicmemory

    An inspectable, correction-aware memory system for AI agents that provides durable context across sessions without coupling to a specific model or framework.

  • Project
    jaredrhod/ai-memory-vault

    A system that turns an Obsidian vault into a persistent, external memory for AI agents, allowing them to remember information across sessions without context limits.

  • Project
    RichmondAlake/memorizz

    A Python framework for building memory-first AI agents that use persistent episodic, semantic, and procedural memory to recall experience and learn new skills over time.

  • Project
    grapeot/context-infrastructure

    A reference implementation and blueprint for building a personalized AI context infrastructure that transforms generic AI responses into personalized assistance through structured memory and user data.

  • Project
    MemoriLabs/Memori

    A persistent, structured memory layer for AI agents that captures conversations and and agent execution traces to eliminate session amnesia and reduce token costs.