← Blog·v0.6.0·

Introducing the Cognitive Cache
The Cache That Thinks.

Every cache stores bytes. cachly v0.6 stores meaning — and now it reasons about it. Five capabilities no cache or AI memory tool has ever had, shipping today.

The problem with “AI memory”

Every AI memory tool today is, at its core, a key-value store with a semantic search layer on top. You call remember(), you call recall(). That's it. Glorified Redis.

None of them ask: Is this memory still true?None of them detect when two memories contradict each other. None of them can answer “why did this break?” by tracing backwards through stored knowledge. And none of them manage themselves — you still have to manually call session_start, learn_from_attempts, session_end, or nothing gets saved.

cachly v0.6 changes all of that.

Five capabilities. None of them exist anywhere else.

🔍

causal_trace — Root cause analysis through memory

When something breaks, you don't just want the fix — you want to know why it broke. causal_trace searches the entire knowledge graph backwards from a symptom to find the root cause chain, what was tried before, and what eventually worked.

$ causal_trace(problem="auth fails after namespace restart")
🔍 Found 7 related memories (3 failures · 2 fixes)
❌ 1. `k8s:namespace-terminating` (relevance 9/10)
→ Cannot create resources in Terminating namespace
❌ 2. `keycloak:jwks-cache-race` (relevance 6/10)
→ Concurrent JWKS map access panics under load
✅ Fix: `k8s:namespace-wait` — PollUntilContextTimeout 3min
🧩 k8s:namespace-terminating → keycloak:jwks-cache-race → current symptom

This is fundamentally different from recall_best_solution. That tool asks “what worked for X?” — causal_trace asks “why is X happening?” Backwards causal reasoning, not forwards lookup.

📊

brain_diff — Git-style diff for AI knowledge

Developers have git log --stat to see what changed in code. Now AI has the same for knowledge. brain_diff shows exactly what was learned, updated, or went stale across any number of sessions.

$ brain_diff(since_sessions=5)
📊 Brain Diff — last 5 sessions
🧠 31 lessons · +12 new · 19 pre-existing
➕ 12 new lessons learned:
  ✅ 🔴 `auth:token-refresh` — refresh at <60s TTL
  ✅ 🟡 `k8s:namespace-terminating` — wait 3min loop
  ❌ `stripe:402-redirect` — 402 blocks fetch()
📚 19 lessons unchanged (stable knowledge base)

knowledge_decay — Temporal confidence scoring

Not all memories are equally trustworthy. A solution stored 140 days ago that was never recalled is far less reliable than one stored last week and recalled 5 times. knowledge_decay makes this visible — every memory gets a confidence score that decays with age and recovers with recall frequency.

$ knowledge_decay()
██████████ 100% — `auth:jwt-refresh` (3d · 5×)
████████░░  80% 🟡 `k8s:limits` (45d · 1×)
█████░░░░░  52% — `deploy:compose` (78d · 0×)
███░░░░░░░  31% 🔴 `stripe:webhook` (112d · 0×)

Decay rates: critical lessons lose 0.2% confidence/day, minor lessons 1%/day. Each recall adds +5% confidence, capped at +25%.

🧬

memory_consolidate — Knowledge distillation

Over time, AI brains accumulate noise: duplicate lessons on the same topic, memories that contradict each other, knowledge that was stored once and never used again. memory_consolidate is garbage collection for your AI's knowledge — it detects contradictions, merges duplicates into one canonical truth, and prunes stale memories that were never recalled.

$ memory_consolidate()
🧬 Knowledge health: 94/100
🔀 `auth`: 4 duplicates → 1 canonical
🔀 `k8s:provisioning`: 3 duplicates → 1 canonical
⚡ `deploy`: success vs failure memories — review manually
✅ Consolidation complete. Knowledge is now sharper.
🤖

autopilot — Zero-config forever memory

This is the endgame. Not AI tools you have to call — an AI that manages its own memory automatically. autopilot generates a CLAUDE.md / .github/copilot-instructions.md that instructs any AI (Claude, GPT-4, Gemini, Cursor, Copilot, Windsurf) to:

  • Auto-recall relevant context at session start
  • Auto-save checkpoints during work
  • Auto-learn from every success and failure
  • Auto-compress memory before context limit
  • Auto-persist everything on session end

One command. Write the file once. Your AI never forgets again — across sessions, reboots, and model switches.

Why this is different

Every other cache / memory tool
Stores bytes. Knows nothing.
All knowledge treated as equally valid forever.
Cannot explain why something broke.
Requires manual session management.
cachly Cognitive Cache
Understands meaning, weights by confidence.
Knowledge decays with age, recovers with recall.
causal_trace traces root cause in seconds.
autopilot manages memory — completely automatically.

Get started

The Cognitive Cache tools ship in @cachly-dev/[email protected]. If you're already set up, run:

$ npx @cachly-dev/mcp-server@latest autopilot

New to cachly? One command configures Claude Code, Cursor, GitHub Copilot, and Windsurf — no dashboard visit, no copy-paste:

$ npx @cachly-dev/mcp-server@latest autopilot

What's next

The Cognitive Cache is the foundation for what comes next: a distributed AI knowledge network where lessons learned by one agent propagate to all connected agents automatically. A cache that doesn't just remember — but one that learns, reasons, and improves over time. We're just getting started.

cachly is a persistent AI Brain for developers — memory shared across Claude Code, Cursor, GitHub Copilot & Windsurf simultaneously. Auto-detects every editor. Bootstraps from your git history. 115 MCP tools. Free tier, EU servers, no credit card.

Your AI is forgetting everything right now.

Every session starts blank. Every bug re-discovered. Every deploy procedure re-explained. cachly fixes that in 30 seconds — your AI remembers every lesson, every fix, every teammate's hard-won knowledge. Forever.

🇪🇺 EU servers · GDPR-compliant🆓 Free tier — forever, no credit card⚡ 30-second setup via npx🔌 Claude Code · Cursor · Copilot · Windsurf
cachly.dev · v0.6.0 · 5. Mai 2026Vollständiger Changelog →