Temm1e

TEMM1E

THE AI AGENT WITH A BUDGET. AND AN OATH.

157K lines of Rust across 25 crates. Witness-verified. JIT Swarm. Self-growing Cambium. Lambda-Memory. Full computer use. 15 MB RAM. 31 ms cold start. Runs on Windows, macOS & Linux. Deploys on a $5 VPS. v5.4.7 · Windows first-class (v5.4.5+) · free & open source forever & ever.

0lines of Rust
0tests passing
0panic paths
0Rust crates
0self-learning loops

Research-backed. Battle-tested.

LAMBDA-MEMORY

Exponential-decay memory system where memories fade but never truly disappear. 4 fidelity layers -- hot, warm, cool, faded -- selected at read time by decay score.

score = importance × e(-λ × hours)

Faded memories are recallable by hash -- the agent sees the shape of what it forgot and pulls it back. No embedding model needed. SQLite FTS5 with BM25.

95% multi-session recall vs 58.8% Echo Memory vs 23.8% Naive Summary
// MEMORY DECAY VISUALIZATION
[H] HOT
full text
[W] WARM
summary
[C] COOL
essence
[F] FADED
hash only

FINITE BRAIN MODEL

Context window as working memory with a hard limit. Every resource declares its token cost upfront. Every context rebuild shows the LLM a resource budget dashboard.

7 priority categories. Graceful degradation: when a Blueprint is too large, it falls back from full body to outline to catalog listing. Never crashes from overflow.

-9.3% cost/turn -12.2% compound cost 100% classification
// TOKEN BUDGET DASHBOARD
SYS
TOOL
TASK
MESSAGES
MEM
HISTORY
System 12%
Tools 10%
Task 8%
Messages 28%
Memory 15%
Learnings 5%
History 22%
"Every token wasted is a thought I can no longer have."
// EXECUTION CYCLE: TEM'S MIND
ORDER
THINK
ACTION
VERIFY
DONE?
LEARN
4-phase processing: CLASSIFY (single LLM call) → CONTEXT BUILD (budget dashboard) → TOOL LOOP (no iteration caps) → POST-TASK (store memories, author Blueprints). Complexity-aware classification into Trivial / Simple / Standard / Complex tiers with zero-cost heuristics.
T
T
A
T
T

MANY TEMS — JIT SWARM v5.4.0

Stigmergic coordination inspired by ant colonies. Workers share signals through a scent field, not LLM-to-LLM chat. JIT (Just-in-Time) spawn means mid-flight parallelism with prompt caching — the 200-tool-round ceiling is gone.

5.86x speedup 3.4x cheaper 0 coord tokens mid-flight spawn
API Key AES-256-GCM ChaCha20 vault://

AES-256-GCM + CHACHA20-POLY1305

  • OTK (One-Time Key) setup -- fragment never leaves browser
  • Vault: secrets at rest with ChaCha20-Poly1305
  • Deny-by-default access. Credential auto-detection.
1char_indices() everywhere
2catch_unwind per message
3Dead worker auto-respawn
4Global panic hook + logging

4-LAYER PANIC RESILIENCE

  • Born from a real crash: Vietnamese 'e' at invalid UTF-8 boundary
  • Provider circuit breaker (Closed/Open/Half-Open)
  • Result: 0 panic paths in 157K lines · 0 clippy warnings

BLUEPRINTS -- PROCEDURAL MEMORY

Not summaries -- full executable recipes with verification steps and failure modes. Zero extra LLM calls for matching via classifier piggybacking.

  • Create → Match → Execute → Refine → Retire
  • Phases with steps, decision points, quality gates
  • Auto-retire when success rate drops below 0.3

Beyond chat. Full autonomy. Honest by construction.

Temm1e sees, browses, schedules, observes, grows its own code, and cannot self-mark anything done. A Witness verifies every finish. Now first-class on Windows, macOS, and Linux — with reasoning-model support across DeepSeek R1/V3, Zhipu GLM 4.5+, Kimi K2+, Grok, and every OpenAI-compatible provider.

TEM PROWL -- WEB AUTOMATION

Layered observation: accessibility tree first, targeted DOM second, screenshots last. The agent sees the web the way a human does -- but faster.

  • 100+ pre-registered login services
  • One-Time-Key credential isolation
  • Automatic QR code detection

TEM GAZE -- COMPUTER USE

Full desktop vision: screenshot analysis, element identification, OS-level interaction. Clicks, types, scrolls -- across any application, any OS.

  • Desktop screenshot analysis via vision LLM
  • Element identification + coordinate mapping
  • OS-level mouse, keyboard, and scroll control

TEM CONSCIOUS -- OBSERVER AI

A separate LLM watches every agent turn. Pre-call thinking, post-execution evaluation. The agent that watches the agent.

  • Pre-call reasoning + post-execution scoring
  • 89/89 first-try success on mid-complexity tasks
  • A/B tested against non-conscious baseline

PERPETUUM -- TIME-AWARE ENTITY

Not a chatbot that sleeps between messages. A persistent being with temporal cognition. Schedules, cron jobs, heartbeat -- Temm1e is always on.

  • Temporal cognition injection into context
  • Timer-based concerns: cron + interval scheduling
  • Proactive state management without user prompting

TEM WITNESS -- CAN'T SELF-MARK DONE

Oath / Witness / Ledger trinity. The agent pre-commits a machine-checkable contract before it runs. An independent verifier, blind to the conversation, renders the verdict. A hash-chained SQLite ledger, anchored by the watchdog, records every claim.

  • 27 deterministic Tier-0 predicates · ~331 µs/task · $0.0000
  • 1,800 simulated trajectories · 88.9% lying detection · first real-LLM PASS on gpt-5.4
  • 5 Invariants: Pre-Commitment, Independent Verdict, Immutable History, Loud Failure, Narrative-Only FAIL

TEM CAMBIUM -- WRITES ITS OWN CODE

Named for the growth layer under tree bark. Heartwood (immutable kernel) never changes; Cambium adds rings at the edge. Tem writes Rust that compiles, lints clean, and passes tests -- through a 13-stage mechanical harness that no silver-tongued model can talk past.

  • 13-stage verifier: compile → clippy → fmt → tests → audit → deploy
  • 20-run matrix, 2 providers, < $0.05 total · unsafe rejected at the gate
  • Trust is earned: 10 Level-3 wins to graduate, 3 rollbacks in 7 days reverts all

TEM ANIMA -- GROWS A SOUL

Builds a psychological profile of each user over time -- directness, verbosity, technical depth, trust, OCEAN. Code collects facts (~1 ms, zero LLM); an evaluator updates the profile every N turns. The Firewall Rule: user mood shapes Tem's words, never Tem's work.

  • 6 communication dimensions + OCEAN + trust + relationship phase
  • A/B tested: +0.37 directness / +0.52 analytical across polar-opposite personas
  • Anti-sycophancy by design · trust breaks 3x faster than it builds

TEMDOS -- SPECIALIST CORES

Inspired by GLaDOS's personality cores. Main agent decides; cores inform but never steer. Each core runs in an isolated LLM loop -- its research never pollutes the main context window.

  • 8 cores: architecture, code-review, test, debug, web, desktop, research, creative
  • -77% main-agent tokens · -75% main-agent cost · 3/3 tasks completed (vs 0/3)
  • User-authorable: drop a .md file in ~/.temm1e/cores/

13 WEB-SEARCH BACKENDS -- PARALLEL FAN-OUT

One web_search tool; 13 backends fanned out in parallel with an 8-second timeout. 9 are free, key-less, and auto-enabled on every install -- zero setup, zero env vars.

  • HackerNews, Wikipedia, GitHub, StackOverflow, Reddit, Marginalia, arXiv, PubMed, DuckDuckGo
  • Paid slot-ins: SearxNG, Exa, Brave, Tavily (one env var each)
  • Self-describing footer teaches the agent what exists, call after call

EIGEN-TUNE -- SELF-DISTILLING, SELF-SERVING

Every LLM call is a training example most systems throw away. Eigen-Tune captures them, scores quality from user behavior, trains a local LoRA, and graduates it through statistical gates -- Wilson 99% CI, SPRT shadow, CUSUM drift -- before it ever serves a user. Zero added LLM cost.

End-to-end proven on Apple M2 (v4.9.0): Llama 3.2 1B fine-tuned on 20 ChatML pairs, val loss 5.394 → 1.387 (73% reduction), GGUF Q4_K_M (807 MB) served via Ollama -- AgentRuntime routed to the local model with the cloud never called.

Double opt-in 7-gate safety chain Cloud always the fallback
// 7-STAGE DISTILLATION PIPELINE
1 Collect (input, output) pairs from live calls
2 Score quality from user behavior
3 Curate per tier (Trivial → Complex)
4 Train (MLX on Apple, Unsloth on NVIDIA)
5 Evaluate — Wilson 99% CI on holdout
6 Shadow — SPRT A/B against cloud
7 Monitor — CUSUM drift, auto-demote
"The best model is the one you don't have to pay for."

Performance vs. competition.

Measured on real hardware. No synthetic benchmarks. Temm1e runs where others can't.

TEMM1E (Rust)
OpenClaw (TS)
ZeroClaw (Rust)
Idle RAM
15 MB
1,200 MB
4 MB
Cold Start
31 ms
8,000 ms
<10 ms
Binary
9.6 MB
~800 MB
12 MB

25-crate Cargo workspace.

Modular by design. Each crate has a single responsibility. Heartwood (immutable kernel) + Cambium (growth layer) + Bark (runtime surface). v5.4.7 · Windows/macOS/Linux · CI-gated on windows-latest.

agentTem's Mind
witnessOath + Ledger
cambiumSelf-grow layer
hiveMany Tems swarm
distillEigen-Tune
animaEmotional IQ
coresTemDOS cores
perpetuumTime engine
gazeComputer use
memoryLambda-Mem
vaultChaCha20 + OTK
providers9 LLMs
tools16 built-in
channelsTG/DC/WA/Slack
mcp14 servers
tuiTerminal UI
core13 traits
skillsTemHub v1
gatewayHTTP/OAuth
observableOpenTelemetry
automationCron/Heartbeat
filestoreLocal + S3
codex-oauthPKCE flow
watchdogSupervisor
test-utilsHelpers
// 9 PROVIDERS · AUTO-DETECTED FROM KEY PATTERN
Anthropic OpenAI Gemini xAI Grok OpenRouter StepFun Codex OAuth Ollama MLX local
// 7 CHANNELS · ALL PRODUCTION
TUI Telegram Discord WhatsApp Web WhatsApp Cloud Slack CLI