predacore 0.1.4

PredaCore — the apex autonomous agent. Hybrid Rust memory kernel, topped LongMemEval R@5 = 0.9574.

The hyper-autonomous AI agent with persistent memory and 55 powerful tools.

0.9574

_{R @ 5   ·   LongMemEval}

Persistent memory. On your laptop. No cloud. No API keys. Yours forever.

---

You:        hey atlas, remember that rate-limiter bug from last month?

PredaCore:  yeah — api_client.py:142, headers dropped on retry. You
            patched it. Similar pattern still lives in webhook_retry.py.
            want me to fix it there too?

Most AI forgets you the moment the tab closes. PredaCore doesn't. Rust kernel, 13 markdown identity files, 55 tools, nine channels, zero vendor SDKs. Every number on this page reproducible in one command.

• Stop re-explaining yourself. It knows your repo, your stack, your architecture — across weeks.
• Bugs don't bite twice. Patterns you debugged last month get flagged when they reappear.
• Preferences stick. Say "use pytest" once — never again.
• Work compounds. Every session picks up where the last left off. Useful memories persist; dead weight fades automatically — preferences live for weeks, casual chats for days, all tuned by session reward.
• You own it. Memory lives in ~/.predacore/. No cloud, no account, no vendor.

You can delete it anytime — rm -rf ~/.predacore/agents/atlas/.

---

🧰 55 powerful tools

Not an LLM wrapper. A digital operator wired into your machine through a hardened dispatcher — Express-style middleware, per-tool circuit breakers, adaptive P95 timeouts, LRU cache, SHA-256-hashed persistent approvals.

	Tools
Code & shell	execute_code (13 langs · sandboxed Docker · optional) · python_exec · run_command · read_file · write_file · list_directory
Web	browser_control (hijacks Chrome via DOM — 100× faster than screenshots) · deep_search · web_search · web_scrape
Desktop / mobile	desktop_control (PyObjC · 1–5ms per action) · screen_vision · android_control (ADB + uiautomator2)
Git	git_semantic_search ("where is the auth middleware?") · git_context · git_diff_summary · git_commit_suggest · git_find_files
Agents & planning	multi_agent — fan-out · pipeline · consensus · supervisor, with optional DAF gRPC process isolation for true parallel agents · strategic_plan (HTN + MCTS, multi-objective) · openclaw_delegate
Memory	memory_store · memory_recall · semantic_search (scoped global · team · scratch)
Identity	identity_read · identity_update · journal_append — writes to the agent's 13-file soul
MCP client	mcp_add · mcp_list · mcp_remove · mcp_restart — mount any MCP server mid-chat
REST APIs	api_add · api_call · api_list · api_remove — bind any service in seconds
Pipelines	tool_pipeline (sequential · parallel · conditionals · templates) · tool_stats
Collective intelligence	skill_evolve · skill_scan · skill_endorse · collective_intelligence_sync · collective_intelligence_status · marketplace_*
Voice / creative / cron	speak · voice_note · image_gen · pdf_reader · diagram (Mermaid) · cron_task
Infrastructure	secret_set · secret_list · channel_configure · channel_install

---

How the engine purrs

Rust compute kernel. Candle BGE + BM25 + trigram fuzzy + entity extraction. SIMD cosine. Deterministic retrieval — no LLM sampling, no RNG. That's why benchmarks reproduce bit-identical.

Thirteen files. One soul. Your agent's identity lives in ~/.predacore/agents/<name>/ as plain markdown. cat them. git log them. rm them. Beliefs graduate observation → working_theory → tested → committed. Every mutation auto-diffs to EVOLUTION.md. Tampered SOUL_SEED aborts startup. Fail closed.

Safety as a primitive. Prompt-injection scan on every identity load. SSRF guard on web tools. Secret-shape allowlist — even yolo can't write arbitrary env vars. Persona-drift regex ladder auto-regenerates drifted turns. Memory scopes (global · agent · team · scratch) prevent cross-contamination.

Per-session lane queue. Same session = serial FIFO. Different sessions = concurrent. Meta-cognition catches loops, oscillation, thrashing — with a diversity exception so real exploration isn't punished.

DAF — true parallel agents. When in-process asyncio isn't enough, the Dynamic Agent Fabric ([server] extra) gives you gRPC multi-process isolation. Agents run in their own processes, crash-isolated, with self-optimization: >20% error rate → respawn · queue depth >10 → scale out · idle >300s → terminate · P95 latency >3× baseline → marked degraded. Wall-clock budgets clamped 10s..6h, hard-killed via asyncio.wait_for. Teams get private 72h-TTL scratchpads so findings don't leak to caller memory.

---

Quickstart

pipx install "predacore[full]"
predacore

One command. First message in under two minutes. Rust ships as pre-built wheels — no toolchain required.

Don't have pipx? brew install pipx (macOS) · python -m pip install --user pipx (Linux/Windows). Already inside a venv or using Conda? Plain pip install "predacore[full]" works too.

Zero-config. The agent configures itself mid-chat:

You: add Anthropic — key sk-ant-api03-XXXXXXXXXX
You: enable telegram with token 123:abc
You: install the GitHub MCP server

Routed through secret_set, channel_configure, mcp_add. Writes land in ~/.predacore/.env (chmod 600).

Install	Adds	Δ Wheel
predacore	Engine · CLI · webchat · 8 channels · Playwright · PDF · voice · sandbox · Rust kernel	~350 MB
predacore[full]	+ spaCy · desktop automation · Android ADB	+200 MB
predacore[server]	+ FastAPI · Redis · Prometheus · DAF gRPC	+150 MB

---

Benchmarks

0.9574 R@5 on LongMemEval — the long-term-memory benchmark from ICLR 2025. 500 conversational histories · ~57M tokens · 470 scored.

Category	n	R@5	R@10	R@20
knowledge-update	72	0.9861	0.9861	1.000
multi-session	121	0.9835	0.9917	1.000
single-session-assistant	56	0.9643	0.9821	0.9821
single-session-user	64	0.9531	0.9844	1.000
temporal-reasoning	127	0.9370	0.9606	0.9843
single-session-preference	30	0.8667	0.9333	1.000

Four of six categories clear 0.95. Bit-identical reproduction:

wget https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_s_cleaned.json
python -m predacore.evals.longmemeval --dataset longmemeval_s_cleaned.json --json-out my_run.json

~55 min on Apple Silicon. Zero per-query API cost. Full artifacts in benchmarks/.

Re-run instantly with response cache: PREDACORE_IDEMPOTENT=1 caches every deterministic LLM call locally (SQLite, 24h TTL). Subsequent benchmark runs skip the API entirely for prompts already seen — useful when quota caps kick in or you want iteration speed on eval tuning. Works with any provider (Anthropic/OpenAI + compat/Gemini).

---

Trust profiles

Profile	Behavior
paranoid	Confirms every tool · ethical keyword guard active
normal (default)	Auto-approves 12 read-only · confirms 16 destructive
yolo	Full autonomy · arg-regex still blocks rm -rf, sudo, dd if=

---

Honest weaknesses (no vaporware)

• Windows desktop operator unimplemented. Networked surfaces work everywhere; desktop_control / screen_vision are macOS + Linux only. Coming soon.
• single-session-preference R@5 = 0.867. Retrieval's weak spot — cross-encoder re-ranker is the planned fix.
• GIL not released in Rust kernel. Concurrent embed() calls serialize. rayon helps within a call.
• yolo has no real cost cap. Arg-regex catches rm -rf, not an obfuscated curl | sh.
• _vendor ships in wheels. Five subpackages bloat the install.

---

Links

Deep dives: Memory · Identity · Tools · Multi-agent · Safety · Autonomy · MCP · Channels · Launch profiles

Issues: github.com/rockshub1/predacore/issues · Security: SECURITY.md · Contributing: CONTRIBUTING.md

---

_{Apache 2.0 · Every claim reproducible from the repo.}