April 9, 2026 The PicoClaw team is releasing v0.2.6
Two big changes in this version: AI finally has memory, and tool execution now runs inside a sandbox
Also includes fixes for the WebSocket connection issues from v0.2.5 and a round of stability improvements
1. Memory System
v0.2.6 adds full memory capabilities — AI no longer starts from zero every conversation
Long-term Memory
AI decides on its own what’s worth remembering — your name, preferences, project context — and writes it to a dedicated memory file
It’s loaded automatically at the start of every conversation, so you never have to repeat yourself
There’s also a daily notes feature: key interaction highlights are recorded each day, and the last 3 days of notes are automatically injected into context
Short-term Memory (Seahorse Engine)
A brand-new context management engine built on SQLite + FTS5 full-text search:
- Smart retrieval: No more dumb truncation — relevant history is searched based on the current topic
- Auto-compression: When conversations get long, LLM-generated summaries preserve key information
- Active recall: AI can proactively search and expand past conversations
- Pluggable design: Implemented via the
ContextManagerinterface — switch with one config line
LOCOMO Benchmark Results
We evaluated memory quality using the LOCOMO academic benchmark (10 real long conversations, ~2,000 Q&A pairs):
| Mode | Evidence Hit Rate | LLM Judge Score |
|---|---|---|
| Traditional truncation | 23.8% | 0.66 |
| Seahorse smart retrieval | 70.0% | 0.84 |
Seahorse’s evidence hit rate is 3× higher than the traditional approach — the advantage is especially clear when information spans multiple conversation turns
2. Process Sandbox Isolation
Process isolation: AI tool execution can now run inside a sandbox environment, no longer operating directly on your system — safer and more reassuring
One line: we put a “safety cage” around AI so it can’t accidentally damage your environment while working
3. Tool & Capability Enhancements
Hooks enhancement: The hook system gains a respond action, plus complete documentation — custom workflows are now more flexible
MCP large text storage: Oversized text results are automatically stored as artifacts — no more blowing up the conversation window
HTTP custom headers: Model provider interfaces support injecting custom headers — easier enterprise integration
write_file improvement: Fixed nested JSON escape semantics and added tests — writing files is more reliable
4. Platforms & Channels
Teams Webhook: New Teams Webhook output channel — another option for enterprise users
Feishu (Lark): Card and file replies now correctly carry context
Model fallback: model_fallbacks now uses an independent provider per candidate — if one fails, the next kicks in automatically
5. Bug Fixes
High-value fixes this release, targeting the core pain points from v0.2.5:
✅ WebSocket connection failure — WebSocket URL now derived from the browser address, no longer depends on backend config, finally resolves the connection nightmare since v0.2.5
✅ WebUI unable to connect to gateway — Fixed gateway connection issue on WebUI startup
✅ Gateway PID handling — Hardened PID liveness checks and WebSocket proxy state management — no more false process detection
✅ Gateway PID file leftovers — PID ownership is now verified and stale files cleaned up — restarts no longer get stuck
✅ Docker startup — Launcher now passes -console and opens networking — running in containers no longer a mystery
✅ Duplicate v in CLI help — picoclaw --help no longer shows vv0.2.5
✅ FreeBSD/ARM compatibility — Seahorse context manager disabled on unsupported platforms — no more crashes
✅ exec terminal control characters — Security-related fix, further hardening planned in upcoming releases
6. Dependency Updates
github.com/pion/rtpupgraded to 1.10.1modernc.org/sqliteupgraded to 1.48.0
Summary
| Category | Update |
|---|---|
| 🧠 Memory system | Long-term memory + Seahorse short-term engine + daily notes |
| 📊 Memory benchmark | LOCOMO: evidence hit rate 70%, Judge Score 0.84 |
| 🆕 Sandbox isolation | Process-level isolation, safer AI execution |
| 🔧 Tool enhancements | Hooks respond, MCP artifact, custom Headers |
| 💬 New channel | Teams Webhook |
| 🐛 Key fixes | WebSocket connection, gateway PID, Docker deployment |
| 🔩 Stability | 10+ bug fixes, FreeBSD/ARM compatibility improved |
This version marks the moment PicoClaw truly learned to “remember” —
Not just a jump in benchmark numbers, but the foundation for a continuous, memory-aware relationship between you and your AI
PicoClaw — Lightweight, cross-platform, fast
Website: picoclaw.io
GitHub: github.com/sipeed/picoclaw
Docs: docs.picoclaw.io
Discord: discord.gg/V4sAZ9XWpN