███████╗██╗ ██╗███╗ ██╗ █████╗ ██████╗ ████████╗██╗ ██████╗ ██╔════╝╚██╗ ██╔╝████╗ ██║██╔══██╗██╔══██╗╚══██╔══╝██║██╔════╝ ███████╗ ╚████╔╝ ██╔██╗ ██║███████║██████╔╝ ██║ ██║██║ ╚════██║ ╚██╔╝ ██║╚██╗██║██╔══██║██╔═══╝ ██║ ██║██║ ███████║ ██║ ██║ ╚████║██║ ██║██║ ██║ ██║╚██████╗ ╚══════╝ ╚═╝ ╚═╝ ╚═══╝╚═╝ ╚═╝╚═╝ ╚═╝ ╚═╝ ╚═════╝
Persistent memory for Claude Code
Cross-session context recallSemantic-first capture & search100% localZero cloud
[ Quick Start ]
Get persistent memory in 30 seconds
[ What It Does ]
Remember Everything
Save decisions, bugs, insights, and context. Everything persists across sessions — nothing gets lost.
Intelligent Search
Multi-pass concept search with fuzzy expansion. Queries split into concepts, each expanded with edit-distance-1 deletions for typo tolerance. BM25 passes fused with vector search via RRF. "fevr" still finds "fever."
Git Intelligence
Auto-index commits, track file co-changes, and extract codebase DNA. Your repo’s history becomes searchable context.
Session Continuity
Five lifecycle hooks — SessionStart, Stop, PreCompact, UserPromptSubmit, PostToolUse — inject context, auto-checkpoint commits, and capture save-intent phrases. Claude picks up exactly where you left off.
Checkpoints
Named, pinnable save-points. /checkpoint <name> mid-session, or say "save progress" and Synaptic's UserPromptSubmit hook catches the intent. Every git commit and plan/spec write auto-checkpoints too. /load <name> injects a saved checkpoint plus its referenced entries back into the conversation.
Smart Recall Ranking
SessionStart ranks handoffs + checkpoints by 0.35 × length + 0.35 × project-match + 0.15 × pinned + 0.15 × recency (3-day half-life). Content shorter than 100 chars is filtered out. No more blind recency ordering — quality wins.
Rules Engine
Define permanent rules enforced every session. Commit-msg hooks block violations, stop hooks scan for breaches. Hard and soft enforcement built in.
Zero Cloud
100% local SQLite with WAL mode. Your data never leaves your machine. No accounts, no subscriptions, no tracking.
Cross-Machine Sync
Sync context across your devices via a private GitHub repo. Append-only, conflict-free, automatic. Full maintenance (decay, promotion, consolidation, smart dedup) runs before every sync push.
Semantic Capture
Every message is classified against semantic anchors — preferences, rules, corrections, debugging patterns. Natural language like "that looks terrible" or "I recommend X" is captured automatically.
Hybrid Fact Extraction
Automatically extracts project facts from tool output — database schemas, API routes, credentials, config. Structural parsing runs instantly; optional Haiku synthesis adds intelligence. Zero config.
Auto-Decay
Three memory tiers — ephemeral (7–21 days), working (21–90 days), longterm (forever). Frequently accessed entries survive longer. Memory stays clean without manual pruning.
Smart Cleanup
Conservative duplicate detection and merging. Subset detection archives entries contained in others; cosine similarity dedup (0.90 threshold) fuses near-duplicates. Content-aware survivor selection keeps the longest entry. Runs automatically during maintenance and before sync.
Smart Auto-Recall
Signal-based context recall at session start. Searches memory when you reference a project, imply continuity, or ask about past decisions. Skips generic questions. No wasted calls, no missed context.
[ Why Synaptic? ]
Claude Code has built-in memory. Here's what changes when you add Synaptic.
CLAUDE.md is static. Auto-Memory saves flat notes with no search.
→ Typed, tagged entries with intelligent multi-pass concept search, fuzzy expansion, and hybrid semantic + keyword retrieval.
No recall of previous conversations. Context lost on restart.
→ Session handoffs, persistent rules, and focus prediction.
Auto-Memory grows forever. No way to prune what’s stale.
→ Three tiers with auto-decay, plus smart dedup that detects subsets and near-duplicates (0.90 cosine threshold). CLI: synaptic cleanup [--dry-run] [--aggressive].
No understanding of your repo’s history or structure.
→ Commit indexing, file co-change detection, codebase DNA.
CLAUDE.md rules are suggestions. Nothing enforces them.
→ Hard enforcement via hooks. Soft enforcement via session injection.
Tool output (DB queries, errors, CLI results) is invisible after the session ends.
→ 3-layer pipeline: structural parsing extracts schemas, routes, and config instantly. Optional Haiku LLM synthesizes structured project facts.
Single machine only. No way to share context.
→ GitHub sync across devices — append-only, conflict-free.
| Synaptic | |
|---|---|
| What it stores | Typed, tagged, tiered entries |
| Search | Multi-pass concept search + fuzzy + RRF fusion |
| Cross-session | Handoffs, chains, history |
| Git awareness | Commits, co-changes, DNA |
| Memory cleanup | Auto-decay + smart dedup + CLI |
| Rule enforcement | Hard + soft enforcement |
| Auto-capture | Semantic anchors + debug patterns |
| Fact extraction | Structural parsing + LLM synthesis |
| Multi-machine | GitHub sync across devices |
Synaptic doesn't replace CLAUDE.md — it complements it. Use CLAUDE.md for static project rules. Use Synaptic for the living knowledge that builds up as you work.
[ How It Works ]
Context Injection
Synaptic injects rules, recent memory, git context, predicted focus, and rule violation warnings. Claude knows what you were doing before you even say a word.
Live + Passive Capture
15 MCP tools for explicit saves + 3 slash commands (/checkpoint, /load, /checkpoints). Every git commit and plan write auto-checkpoints; save-intent phrases like "save progress" fire UserPromptSubmit. Semantic transcript scanning classifies messages against 6 concept anchors — no manual effort required.
Auto-Distill
Semantic capture across all messages, hybrid fact extraction from tool output (structural parsing + LLM synthesis), directive detection, debugging pattern extraction, handoff summaries, and rule violation checks.
[ Under The Hood ]
┌────────────────────────────────┐ │ SYNAPTIC v1.7.2 │ ├────────────────────────────────┤ │ │ │ ┌──────────┐ ┌──────────┐ │ │ │MCP Server │─▶│ SQLite │ │ │ │ 14 Tools │ │ WAL Mode │ │ │ └──────────┘ └──────────┘ │ │ │ │ │ │ ▼ ▼ │ │ ┌──────────┐ ┌──────────┐ │ │ │ 3 Hooks │ │FTS5 Index│ │ │ │ Lifecycle │ │+Vec (AI) │ │ │ └──────────┘ └──────────┘ │ │ │ │ │ │ ▼ ▼ │ │ ┌──────────┐ ┌──────────┐ │ │ │ Embedder │ │ Hybrid │ │ │ │ MiniLM-L6 │ │ Search │ │ │ └──────────┘ └──────────┘ │ │ │ │ │ │ ▼ ▼ │ │ ┌──────────┐ ┌──────────┐ │ │ │ Semantic │ │ Signal │ │ │ │ Anchors │ │ Boost │ │ │ └──────────┘ └──────────┘ │ │ │ │ │ ▼ │ │ ┌──────────┐ │ │ │ GitHub │ │ │ │ Sync │ │ │ └──────────┘ │ │ │ └────────────────────────────────┘