Synapse

v2026.05.1-beta.0▼
Search documentation...⌘K
GitHub Star
BETA You are viewing the documentation for the upcoming v2026.05.1-beta.0 release.

Persistent Memory

Synapse provides persistent project memory that survives across sessions, projects, and conversations. Unlike standard context windows that are wiped clean when a chat ends, Synapse memory persists in a local SQLite database, allowing AI agents to build a cumulative understanding of your project.

ℹ️The Statelessness Problem

Modern AI models have massive context windows, but they are still stateless. Every new chat is a "blank slate." Synapse breaks this cycle by providing a "hard drive" for AI reasoning.

Key Features

1. Teacher-Mode Instructions (teach)

The teach tool is the primary way to shape an agent's long-term behavior. Use it to record team-wide coding conventions, architectural constraints, or "never do X" rules.

  • Persistence: These rules are automatically surfaced via agent_prime in future sessions.
  • Priority: Teacher-mode memories carry higher weight in retrieval.

2. Outcome Capture (capture_outcome)

The "Winner" state of a task is often the most valuable piece of context for the next developer. capture_outcome allows an agent to record:

  • What was achieved.
  • Which files were changed and why.
  • Any new "Gotchas" discovered during the process.

3. Semantic Deduplication

To prevent "context bloat," Synapse uses Vector Similarity Analysis (via sqlite-vec). When a new memory is stored:

  • Synapse calculates its cosine similarity against existing records.
  • If a near-identical fact exists, the agent is notified to update the existing memory instead of creating a duplicate.

4. Scoped Isolation (Nests)

Memories can be isolated into Nests. This is critical for developers working on multiple client projects.

  • Context from "Project A" will never bleed into "Project B" unless explicitly requested.

Memory Tool Suite

ToolCapabilityUse Case
memory_store Semantic Persistence Save a design decision or lesson learned.
memory_recall Vector Search Ask "What did we decide about the API?"
teach Rule Ingestion Record team-wide coding conventions.
capture_outcome Outcome Tracking Log the result of a completed milestone.
agent_prime Total Rehydration Get all relevant lessons at the start of a task.

Real-World Workflow: The Learning Loop

  1. Start: Agent calls agent_prime and learns that "We use Vitest for this repo."
  2. Execute: Agent writes a test using Vitest.
  3. Learn: Agent discovers that "Vitest needs jsdom for this specific module."
  4. Store: Agent calls memory_store to record the jsdom requirement.
  5. Close: Agent calls capture_outcome to summarize the new test coverage.
  6. Future: Next week, a different agent works on the same module and automatically receives the jsdom tip.

Next: See how memories connect to structured architectural facts in the Knowledge Graph.