Temm1e

TEMM1E

THE AI AGENT WITH A BUDGET.

Autonomous AI runtime in Rust. Lambda-Memory. Swarm intelligence. AES-256 encryption. 15 MB RAM. 31 ms cold start. Deploys on a $5 VPS.

FUND TEMM1E VIEW SOURCE
0lines of Rust
0tests passing
0panic paths
0Rust crates
0LLM providers
// CORE TECHNOLOGY

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
95.0% 5-session recall
0 embedding models
BM25 on FTS5

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 -- SWARM INTELLIGENCE

Stigmergic coordination inspired by ant colonies. Workers share signals through a scent field, not LLM-to-LLM chat. Zero coordination tokens.

5.86x speedup 3.4x cheaper 0 coord tokens
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 73K lines

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
// BENCHMARKS

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
// ARCHITECTURE

17-crate Cargo workspace.

Modular by design. Each crate has a single responsibility.

agentTem's Mind
hiveSwarm AI
memoryLambda-Mem
vaultEncryption
providers8 LLMs
tools14 built-in
channelsTG/DC/Slack
mcp14 servers
tuiTerminal UI
core13 traits
skillsTemHub v1
gatewayHTTP/OAuth
observableOpenTelemetry
automationCron/Heartbeat
filestoreLocal + S3
codex-oauthPKCE flow
test-utilsHelpers
DONATE