Skip to main content
Glama
Nas01010101

mnemo

by Nas01010101
OverviewSchemaRelated ServersScoreDiscussions
Python
Remote

Tenet

Agent Memory as a Self-Consistent World Model

paper license Qwen Cloud MCP

A memory that stays true as the world changes. Built for the Global AI Hackathon with Qwen Cloud — Track 1.

LLM-agent memory is almost always retrieval over a log of past turns. That's the wrong abstraction for an agent modeling a changing world: as a fact is updated over a long interaction — knowledge churn — stale versions crowd the retrieval budget and the agent answers with an out-of-date value. Tenet reframes memory as a self-consistent belief state — a compact world model of the user — and stays correct where retrieval collapses.

knowledge churn

As a fact is updated 2→12 times, RAG-memory falls 100%→50%. Tenet holds 100%.

Why it's different

retrieval memory (RAG)

Tenet

abstraction

document index of turns

belief state (world model)

a changed fact

two similar passages

superseded (bi-temporal, history kept)

stale evidence

retrieved forever

retired (belief–evidence consistency)

write policy

store everything

surprise-gated (predictive coding)

forgetting

none (grows forever)

salience-decay sweep

queryable across time

no

time-travel (recall(as_of=t))

read path

—

no LLM call

Read the 2-page paper: paper/tenet.md.

Related MCP server: sostenuto

Results (LongMemEval_S, n=40, gpt-4o reader — honest, reproducible; detail in docs/BENCHMARK.md)

recall@10

QA acc

reader tokens

acc / 1k tok

full-context

—

65%

~124,000

0.5

RAG

95%

65%

2,101

30.9

Tenet

97.5%

52.5%

1,067

49.2 ← best

  • Best accuracy-per-token (1.6× RAG; half its context) — and reader-robust: with a frontier reader (claude-opus-4.8) it's 1.7× (Tenet 53.9 vs RAG 32.1).

  • Churn-robust: 100% at every update level while RAG collapses to 50% — and the collapse holds under a gpt-4o reader, so it's structural, not reader weakness.

  • Ablation: the belief–evidence consistency rule alone lifts current-value accuracy 55%→100%.

  • Honest: a strong RAG wins raw one-shot accuracy (65 vs 52.5); Tenet's weak spot is multi-session synthesis. We report it. (Eval off-Qwen; shipped system uses Qwen Cloud.)

The agent

Tenet ships as a personal assistant (src/agent.py) on Qwen Cloud:

you › Hi! I'm Wissem, I live in Montreal and work as a data analyst.
assistant › Nice to meet you, Wissem! How's the analyst work in Montreal?   [remembered 2 facts]
… weeks later …
you › I moved to Toronto and got promoted to senior analyst!
you › Where do I live and what's my job now?
assistant › You live in Toronto and you're a senior analyst. Congrats on the promotion!
python src/agent.py            # interactive assistant
python scripts/demo_agent.py   # the scripted story (video walkthrough)

Quickstart

cp .env.example .env && chmod 600 .env      # add DASHSCOPE_API_KEY (Qwen Cloud)
pip install -r requirements.txt
python scripts/smoke_test.py                # verify connectivity
uvicorn api:app --host 0.0.0.0 --port 8000  # (from src/) HTTP API incl. POST /chat
python src/mcp_server.py                     # or the MCP server (learn/recall/forget/stats)

Reproduce the paper

python scripts/test_memory.py ; python scripts/test_tenet_e2e.py               # capabilities
python scripts/bench_horizon.py --principals 12 --k 6 --updates 2,4,6,8,10,12  # Fig. 1 (churn)
python scripts/lme_recall.py --limit 20 --k 10 --qa --seed 2                    # Table 1 (frontier)
python scripts/bench_knowledge_update.py --principals 4                         # ablation + efficiency
# off-Qwen: prefix with  LLM_PROVIDER=openrouter EMBED_PROVIDER=local OPENROUTER_MODEL=openai/gpt-4o-mini

Architecture

architecture

Two layers over one bi-temporal store (beliefs + evidence), two surfaces (MCP + HTTP), powered by Qwen Cloud (Alibaba Cloud Model Studio). Details: docs/DESIGN.md, positioning vs Mem0/Zep/Letta/Mastra: docs/COMPARISON.md.

Repository

paper/tenet.md            the paper
src/  agent.py            the assistant
      tenet.py memory.py distill.py config.py   the belief-state memory engine
      mcp_server.py api.py alicloud_oss.py       surfaces + Alibaba Cloud deploy
scripts/ demo_agent.py    video walkthrough
         bench_horizon.py bench_knowledge_update.py lme_recall.py   benchmarks
         test_memory.py test_tenet_e2e.py smoke_test.py            tests
docs/ BENCHMARK.md COMPARISON.md DESIGN.md DEPLOY.md SOTA.md  architecture.svg horizon.svg

Citation

@misc{tenet2026,
  title  = {Tenet: Agent Memory as a Self-Consistent World Model},
  author = {Anas},
  year   = {2026},
  note   = {Global AI Hackathon with Qwen Cloud, Track 1},
  url    = {https://github.com/Nas01010101/tenet}
}

License

MIT — see LICENSE.

This server cannot be installed

A
license - permissive license
-
quality - not tested
B
maintenance

How are these scores calculated?

Maintenance

–Maintainers
–Response time
–Release cycle
–Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/Nas01010101/tenet'

If you have feedback or need assistance with the MCP directory API, please join our Discord server