TIMPs

Dual-mode search combines SQL keyword matching (ILIKE) with Qdrant vector similarity retrieval. TIMPs does not rely on a single search method; instead, it runs both an ILIKE query on PostgreSQL and a cosine similarity search on Qdrant vector store simultaneously. This ensures that even if the exact keyword is not present, semantic meaning can still retrieve relevant memories. For example, searching for "React" using the `!blame` command will match both the explicit fact "React + TypeScript + Zustand" via SQL and any semantically similar reflections. The results are presented with importance scores and retrieval counts, allowing users to gauge relevance. This dual approach dramatically improves recall accuracy compared to traditional keyword-only or vector-only systems, especially when memory terms vary or vague queries are used.

Project isolation, ephemeral mode, and model agnostic support are additional key capabilities. Project isolation scopes memories per project, preventing cross-contamination between different codebases, clients, or personal contexts. This is critical for developers juggling multiple projects—each project maintains its own memory namespace, like having separate assistant profiles. Ephemeral mode (`--mode ephemeral`) allows users to run a session that leaves zero persistent trace; nothing is stored beyond RAM, suitable for sensitive discussions or one-off queries. TIMPs is also model agnostic: users can switch between OpenAI GPT-4, Google Gemini, or local Ollama on the fly using the `--provider` flag. This flexibility means users are never locked into a single AI provider and can optimize for cost, performance, or privacy as needed.

TIMPs follows a structured agent architecture with three core components: Planner, Executor, and Reflection. When a user sends a message, the Planner breaks down the goal into steps, the Executor runs tools (file, web, search), and after each turn, the Reflection component extracts and scores memories. Memories are stored in a tiered system: ShortTermStore manages the rolling context window with a token budget, LongTermStore persists facts in PostgreSQL with the 14-field schema and ILIKE search, and VectorStore stores embeddings in Qdrant for similarity search. Commands like `!blame`, `!forget`, `!audit` allow users to manually search, delete, or review memories. This multi-layered approach ensures relevant context is always available while keeping the conversation focused and token-efficient.

Concrete use cases demonstrate TIMPs' value. A developer working on multiple projects can remember each project's tech stack, preferences, and goals without manual notes. For instance, after telling TIMPs "I'm working on a React app with TypeScript and Zustand", the agent stores this as a high-importance reflection. Later, when asking for coding suggestions, TIMPs retrieves the stored context automatically. Another case is using ephemeral mode for sensitive debugging sessions where no memory should persist—ideal for handling confidential client data. Researchers can leverage the `!blame` command with dual-mode search to locate past findings across hundreds of conversations. The outcome is significantly reduced cognitive load and faster, more informed AI assistance.

Target users include software developers, DevOps engineers, open-source contributors, and AI enthusiasts who need context-aware assistance across long-running projects. Platform support covers terminal (TUI/CLI), web browser (via REST API), and desktop app. The tech stack relies on PostgreSQL for structured memory, Qdrant for vector storage, and supports Ollama, OpenAI, and Gemini as model providers. TIMPs is released under the MIT license and is completely free to use—no subscriptions or hidden costs. For those who prioritize data sovereignty, running fully local with Ollama ensures zero data leaves the machine. This combination of persistent memory, privacy, and flexibility makes TIMPs a compelling choice for anyone who wants their AI to truly know them and evolve with every conversation.