mnemon-memory-agent: Long-Term Memory with Read-Time Judgment for DeepSeek Harness
grivn/mnemon-memory-agent
Long-term memory for DeepSeek Harness: raw dated records judged fast by a System 1 model and answered by System 2 from under 4k tokens of context.
mnemon-memory-agent is a long-term memory plugin that runs inside DeepSeek Harness, keeping conversations as raw dated records and doing its retrieval and judgment only when a question arrives. It targets the burden of write-time extraction, where facts, entities and preferences must be defined in advance and every new kind of data needs another set of rules, by moving that work to read time. A fast decision model handles the small yes/no judgments of System 1, an LLM handles query writing and answer composition as System 2, and a background consolidation pass turns records into an overlay that points back to them. It runs as a second instance beside the main agent, leaves that agent unchanged, and publishes one View per turn.
How to Install
dsh plugin --profile web add github:grivn/mnemon-memory-agent- Category
- Data & Knowledge
- Platform
- DSH Plugin
- Author
- grivn
- Distribution
- Plugin
mnemon-memory-agent Key Features
mnemon-memory-agent Repo Summary
What Does It Do?
Mnemon is a DSH plugin for DeepSeek Harness that gives LLM assistants long-term memory by keeping conversations as raw, dated records and doing its work only when a question arrives. It solves the problem of write-time extraction, where a system must decide in advance what counts as a fact, an entity or a preference and needs a new schema for every new kind of data. It borrows from dual-process accounts of thinking: a fast decision model acts as System 1, making many small yes/no judgments about the records searches return, while an LLM acts as System 2, writing a few search queries, naming what the reply needs and composing the answer. It runs as a second instance of DeepSeek Harness beside the main agent, which it leaves unchanged, and publishes one View per turn.
Core Features
- Raw records with read-time judgment: nothing is decided about a record when it is written, so the store only needs a search route that returns dated records.
- Dual-process split: System 1 handles small independent judgments such as whether the reply should use a record or whether it is no longer current, and System 2 is called only when no judged record satisfies a need.
- Consolidated index overlay: a background pass consolidates each record once into topic timelines, value histories and standing instructions that point back to the records, so questions about a whole conversation reach evidence their own searches miss.
- Low context cost: about 3.8k tokens are sent to the answering model per question, and it is the only system above 80% on both LoCoMo and LongMemEval-S below 4k tokens.
- Bounded cost at scale: from BEAM-100K to BEAM-10M, with 80 times as many records, cost per question grows by a factor of 1.11 and critical-path work stays about the same.
- Reproducible research: the repository ships the paper, results and reproduction material so the benchmark numbers can be checked inside the harness.
How to Use This Plugin?
After enabling it in DSH, it runs as a second instance alongside the main agent, which stays unchanged, so you simply keep talking and memory is handled per turn. When a question arrives it first has System 1 judge the retrieved records, then lets System 2 compose the answer, and it publishes one View per turn so you can inspect what happened. To adjust behavior, switch the memory sources, strategies and workspace modules in its configuration; to verify results, follow the repository's LoCoMo, LongMemEval-S and BEAM reproduction flow. The plugin is maintained by Grivn under the MIT license, written in TypeScript and last updated on 2026-09-28.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: grivn/mnemon-memory-agent. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.
DSH Plugin FAQ
