How does the router decide between notes and raw history?

The router judges whether retrieved notes contain enough evidence for this specific question; a short question can still force raw retrieval when it needs an exact detail the summary dropped.

Why keep both compact notes and full raw pages?

Summaries give fast recall while linked raw pages preserve exact constraints, timestamps, and observations that future questions may need.

How is the Jev variant different from the main TierMem setup?

The Jev variant replaces router training and external stores with Markdown notes, searchable history, and Jev routing, runnable as an agent, a Python API, or MCP.

Does recovered evidence improve later queries or vanish?

Recovered facts can be consolidated back into compact memory with source links kept, and the paper studies this across evaluation epochs rather than as a single write.

What are the prerequisites for the full TierMem stack?

Python 3.10 or higher, a running Qdrant vector database, an API endpoint for answer generation and embeddings, a served router for the default LoCoMo and LongMemEval configurations, and a CUDA-capable GPU for router training.

Where does the trained router come from, and how is it served?

The router is published on Hugging Face as FreedomIntelligence/TierMem and decides whether a query can be answered from summaries or needs raw evidence. The LoCoMo and LongMemEval runners expect a vLLM router at http://localhost:8000/v1 naming Qwen3-0.6B, adjustable with --router-base-url and --router-model.

Do I need a vector database for the Jev variant of TierMem?

No. The Jev TierMem variant needs no router training, mem0 or vector database; it uses Markdown notes plus searchable original messages and tool output, and Jev decides when to look beyond the notes. Recovered facts can be written back when they are supported and useful.

CommunityJev Agents & MCP

Tiermem

Tiered agent memory keeping compact summaries with source links over searchable raw history, where a router decides per query whether notes suffice or raw pages must be recovered.

See highlights

Open github.com

Stars: 14
Language: Python
License: MIT
Published by: Community
Author: FreedomIntelligence
Last verified: Oct 2026

FreedomIntelligence/Tiermem

Organization repository on GitHub

View on GitHub

[COLM' 2026] TierMem: Balancing Compressed Memory and Raw Evidence for Long-Horizon Agent Memory

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Qiming Zhu · Shunian Chen · Rui Yu · Zhehao Wu · Benyou Wang. Think fast with summaries. Dig deep into raw history. Paper / OpenReview · Earlier preprint · Model · Citation. Use cases · Benchmarks · Quick start · Training. 2026-09-22 — Try Jev TierMem: lightweight agent memory with Markdown notes, searchable raw history, and Jev routing. Run the agent, use the Python API, or connect through MCP.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Compact tier answers fast; raw tier recovers exact details through bounded extra retrieval.
Every compact entry keeps links to the source pages that support it.
Jev variant needs no router training, external memory service, or vector database.
Ships runners recording answer quality, routing, token use, and latency on memory benchmarks.

Watch out

Needs a vector database, an answer-generation endpoint, and a served router for the default benchmark configs, plus GPU time for retraining. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

liteagents

Python SDK that runs the same prompt, tools, and MCP setup across DeepAgents, Pydantic AI, Claude Agent SDK, Codex, and OpenCode by switching harness.