Your AI agent is brilliant — until its context fills up and it forgets everything like Dory. AgentSynapse gives it a memory that actually sticks, across sessions, restarts, and context compressions.
Get Started Free →AgentSynapse captures every tool call, stores semantic embeddings, and makes past context searchable — so your agents carry knowledge forward instead of starting from zero every session.
PostToolUse hooks fire after every agent tool call, sending compact summaries to AgentSynapse without blocking the agent. Noisy housekeeping tools (TodoRead, TodoWrite, TaskList, TaskGet) are filtered out automatically — only signal gets stored, not noise.
Combines pgvector cosine similarity with PostgreSQL full-text search (websearch_to_tsquery English) using Reciprocal Rank Fusion. A configurable decay factor weights recent memories higher, and access counts let the system detect and flag stale entries over time.
Automatically extracts typed entities — services, databases, providers, files, concepts, technologies — and links them with typed relationships: uses, hosted_on, depends_on, caused_by, deployed_with. Ask what breaks if you change something and get a real traversal, not a guess.
Built-in project planning with dependency-aware task graphs: create_spec, add_task, complete_task, get_ready_tasks, update_spec_status, get_active_spec, link_memory_to_task. After context compression, get_active_spec reloads the full task graph in one call.
Raw captures are processed through Bronze, Silver, and Gold stages: summarization, deduplication at cosine similarity threshold 0.92, and entity extraction — all running asynchronously with a semaphore-limited concurrency of 8. Superseded chunks are marked, not deleted.
20+ MCP tools across three categories: Memory Management (store_memory, search_memories, get_recent, get_by_file, list_sessions, get_session_context, merge_memories, get_memory_health), Entity and Relationships (get_project_graph, get_project_facts, get_entity_relationships, find_dependents, find_path, get_entity_history), and the Spec Planning Engine — all available natively in Claude Code.
AgentSynapse runs as a local background Python service — no cloud account, no data leaving your machine. PostgreSQL with pgvector stores 768-dimensional embeddings generated by Ollama locally. If Ollama is unavailable, search gracefully degrades to full-text only.
AgentSynapse is the memory layer. Every tool in the suite can write to it and read from it — building shared context across your entire agent workflow.