mnemon-ai 1.1.0

Cut LLM agent token costs by 93%. Execution cache for LangChain, CrewAI, AutoGen, LangGraph — zero tokens on repeat runs, 2.66ms latency vs 20s.

Mnemon

Stop paying for work your agent already did. Watch it get better every run.

Quickstart · Benchmarks · How it works · Install · API

---

Mnemon gives your agent two things:

Token and latency savings — repeated tasks skip the LLM entirely. 2.66ms instead of 20 seconds. Zero tokens instead of 1,250. The more your agent runs, the more it saves.

A learning loop — every outcome is observed, every pattern is detected, every failure is quarantined. Your agent doesn't just cache work — it accumulates intelligence that makes the next run cheaper and better than the last.

One line of code. No infrastructure. No changes to your existing agent.

Built for: agentic workflows that repeat · production AI agents with high API costs · teams optimizing LLM infrastructure spend · LangChain, CrewAI, AutoGen, and LangGraph pipelines running at scale.

import mnemon
m = mnemon.init()

---

The Problem

Every agent framework — LangChain, CrewAI, AutoGen, LangGraph — is stateless by default.

Your agent generates a security report for Acme Corp every Monday. Every Monday it starts from zero: re-reads the same context, re-reasons through the same structure, re-generates the same plan. You pay full LLM price each time. It never gets faster. It never gets smarter.

This is the core problem with agentic workflow cost optimization: every run is treated as the first run. Token usage doesn't decrease. API costs don't decrease. Latency doesn't decrease. You scale your agent, you scale your bill.

You built a smart agent. You got an amnesiac that invoices you twice.

---

Quickstart

Path 1 — zero code changes (recommended)

Already using Anthropic SDK, OpenAI, LangChain, LangGraph, CrewAI, or AutoGen? Add two lines. Everything else stays the same.

import mnemon
mnemon.init()   # auto-detects installed frameworks and patches them

# your existing code — completely unchanged
from anthropic import Anthropic
client = Anthropic()
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Generate weekly security report for Acme Corp"}]
)

Mnemon intercepts the call, checks its cache, and returns a cached response instantly on repeat runs — no API call made, no tokens spent.

Path 2 — explicit caching with m.run()

For any expensive recurring computation that isn't a direct framework call. generation_fn is your real logic — only called on a cache miss.

import mnemon
from anthropic import Anthropic

client = Anthropic()
m = mnemon.init()

def generate_report(goal, inputs, context, capabilities, constraints):
    # only runs on a cache miss — put your real LLM call here
    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=[{"role": "user", "content": goal}],
    )
    return response.content[0].text

result = m.run(
    goal="weekly security audit for Acme Corp",
    inputs={"client": "Acme Corp", "week": "Apr 21-25"},
    generation_fn=generate_report,
)

print(result["output"])            # your actual result — the return value of generation_fn
print(result["cache_level"])       # "system1" | "system2" | "miss"
print(result["tokens_saved"])      # 1250 on a cache hit, 0 on first run
print(result["latency_saved_ms"])  # 20000.0 on a cache hit

generation_fn also accepts async def — both work.

See what you've saved

m = mnemon.get()             # retrieve the running instance from anywhere
print(m.waste_report)        # repeated queries and their cumulative cost
print(m.get_stats())         # EME stats, bus signals, DB stats

---

The Numbers

Execution cache — EME benchmarks

System 1 hit (exact match)	2.66ms
Fresh LLM generation	~20,000ms
Speedup	7,500×
50 concurrent agents, burst	0 LLM calls · 0.18s total
Tokens saved (50 agents)	62,500
Cost saved (50 agents)	$0.94

At scale (80% System 1 + 15% System 2 hit rate)

Daily plans	Monthly cost saved
100	$56
1,000	$503
10,000	$5,034
100,000	$50,344

What your session looks like

First run (cache miss — plan is stored):

Mnemon: 1 plan(s) cached → next run saves ~1,250 tokens (~$0.0038)

Every run after (cache hit):

Mnemon: ~1,250 tokens saved · ~$0.0038 · 20.0s faster

Full runs, methodology, and raw data: reports/

Academic validation: Stanford researchers published Agentic Plan Caching: Test-Time Memory for Fast and Cost-Efficient LLM Agents at NeurIPS 2025, measuring 50.31% cost reduction with the same approach. Mnemon is the production implementation of this idea — one import, works today.

---

Two Things That Fix This

1. Execution Memory Engine — save tokens and time

The EME is a generalised execution cache for any expensive recurring computation. After the first run, Mnemon fingerprints the plan and stores it. Every subsequent run with the same — or semantically similar — goal is served from cache.

First run:  20,000ms · 1,250 tokens · full cost
Every repeat:  2.66ms · 0 tokens   · $0.00

It works in two modes:

• System 1 — exact fingerprint match. Sub-millisecond. Zero LLM calls.
• System 2 — semantic similarity match. "Weekly security report" hits the cache for "generate security audit". Only changed segments go to the LLM. Requires pip install mnemon-ai[full] or OPENAI_API_KEY in your environment.

Failed segments are quarantined by the Retrospector — bad patterns can't recycle into future plans.

Ships with 100 pre-warmed segments from real enterprise runs so the cache starts warm on day one.

2. Experience Bus — a learning loop that never stops

The Bus is a passive observer. Every computation outcome — success, failure, latency, pattern — is recorded and analysed in the background. You never call it directly. It's always running.

What it detects:

Signal	What it means
DEGRADATION	latency spike vs rolling baseline
PATTERN_FOUND	a task type is failing at >30% — before you notice
ANOMALY	sudden failure after a string of successes
RECOVERY	the agent is stable again after a failure streak

What it does with that intelligence: feeds it back to the EME. Success patterns strengthen the fragment library. Failure patterns trigger quarantine. The cache gets smarter on every run — not just bigger.

This is the loop: EME saves the work. Bus learns from it. Both get better.

---

Zero Code Changes

Mnemon patches your installed frameworks at the call level. One import, nothing else:

import mnemon
m = mnemon.init()

# everything below is unchanged
from langchain_anthropic import ChatAnthropic
llm = ChatAnthropic(model="claude-sonnet-4-6")
response = llm.invoke("Generate weekly security report for Acme Corp")

Supported frameworks:

Framework	What gets patched
Anthropic SDK	client.messages.create
OpenAI SDK	client.chat.completions.create
LangChain	BaseChatModel.invoke / ainvoke
LangGraph	CompiledGraph.invoke / ainvoke
CrewAI	crew kickoff via event bus hook
AutoGen	ConversableAgent.generate_reply

Framework notes:

• LangGraph — call mnemon.init() before compiling your graph.
• CrewAI — import crewai before calling mnemon.init().

---

vs. Everything Else

	Mnemon	Mem0	LangMem	Roll your own
Execution caching (skip LLM entirely)	✅	❌	❌	❌
System learning loop	✅	❌	❌	❌
Zero-code auto-instrumentation	✅	❌	❌	❌
Runs fully local (no cloud, no API)	✅	❌	❌	✅
Drift detection	✅	❌	❌	❌
Multi-tenant isolation	✅	✅	❌	⚠️
One-line setup	✅	❌	❌	❌

Every other library makes your prompt slightly better. Mnemon eliminates the LLM call on repeated work and makes the next run cheaper than the last.

---

Install

pip install mnemon-ai

No API key needed to start:

mnemon demo     # see it working in 30 seconds
mnemon doctor   # health check

System 2 semantic recall

System 2 matches semantically similar goals — "weekly security report" hits the cache for "generate security audit". Two ways to enable it:

# Option 1 — offline, no API key (recommended for production)
pip install mnemon-ai[full]

# Option 2 — if you already have an OpenAI key in your environment
export OPENAI_API_KEY=sk-...
# that's it — Mnemon detects it and activates System 2 automatically

Without either, Mnemon runs System 1 only (exact-match cache) and prints a one-line message telling you what to do. System 1 is still valuable — just no semantic matching.

pip install mnemon-ai[full]         # sentence-transformers (offline, 85% precision)
# or set OPENAI_API_KEY            # uses text-embedding-3-small (90% precision)

---

API

Init

m = mnemon.init()                             # global singleton
m = mnemon.init(tenant_id="acme_corp")        # explicit tenant
m = mnemon.init(silent=True)                  # suppress session summary
m = mnemon.init(eme_enabled=False)            # bus + MOTH only
m = mnemon.init(bus_enabled=False)            # EME + MOTH only
m = mnemon.get()                              # retrieve the running instance

Diagnostics

report = m.drift_report()   # cross-session degradation analysis
stats  = m.get_stats()      # EME, bus, watchdog, DB stats
print(m.waste_report)       # repeated queries + cost

CLI:

mnemon doctor   # health check
mnemon demo     # live demo

Async

from mnemon import Mnemon

async with Mnemon(tenant_id="my_company") as m:
    result = await m.run(
        goal="weekly security audit for Acme Corp",
        inputs={"client": "Acme Corp", "week": "Apr 21-25"},
        generation_fn=my_planning_function,
    )
    print(result["output"])         # your actual result
    print(result["cache_level"])
    print(result["tokens_saved"])

Production — multi-tenant

from mnemon import Mnemon
from mnemon.security.manager import TenantSecurityConfig

m = Mnemon(
    tenant_id="acme_corp",
    security_config=TenantSecurityConfig(
        tenant_id="acme_corp",
        blocked_categories=["pii", "medical_records"],
        encrypt_privileged=True,
    ),
    enable_watchdog=True,
    enable_telemetry=True,
)

Each tenant_id gets an isolated SQLite database — no cross-tenant leakage.

From config

m = Mnemon.from_config("./mnemon.config.json")

---

Fail-Safe

Mnemon never crashes the system it wraps.

What fails	What happens
EME cache	generation_fn called directly
Experience bus	agent continues unmonitored
Database unavailable	in-memory fallback

All failures are logged, never raised. You can't break your agent by adding Mnemon.

---

Why This Exists

These aren't hypothetical — we filed these issues before writing a line of Mnemon:

• CrewAI #4415 — context pollution and DB write contention in multi-agent runs
• Dify #32306 — redundant reasoning tax in agent nodes
• Kimi CLI #1058 — context saturation in 100-agent swarms
• E2B #1207 — environmental amnesia across sandbox restarts

---

License

MIT. Free to use, free to build on.

---

_{Mnemon was Alexander the Great's personal historian — the one whose only job was to ensure nothing was ever forgotten, so every campaign built on the total accumulated knowledge of every campaign before it.
Your agents have a Mnemon now.}