· Research software
TOMC Memory
Task-Oriented Memory Compilation
Open-source context compression and task memory for Claude Code and Codex.
汤姆 C：为 AI 助手整理、压缩上下文，保存任务进度。
Overview
TOMC Memory prepares context for a task from a long history. It computes supported state, relation, and count records, then combines those records with selected source text under a memory budget. The resulting plain-text context can be used by the same language model that answers the task.
Construction runs on CPU without training or an auxiliary neural model. The software includes a Python API and an assistant plugin, with optional task notes that can be explicitly saved and recalled in a later session.
Watch the demo
From temporal representation to compact histories
TOMC Memory builds on research in temporal representation: representing a sequence compactly while retaining information useful for a task. Time-series methods such as piecewise aggregate approximation (PAA) and Symbolic Aggregate approXimation (SAX) provide the motivation.
TOMC explores this representation question in conversational histories. Its task-oriented records make selected information explicit, while retained source text provides supporting detail. Sequence compression and symbolic representation are the conceptual link to time-series representation; TOMC uses its own record-construction procedure.
How it works
- 1. Provide the task
Supply the history, the next question or task, and a memory budget.
- 2. Construct the representation
Compute supported task records and select relevant source passages.
- 3. Use the context
Pass the prepared text to the reader, or recall explicitly saved notes for the next session.
For ongoing work, saved notes can contain decisions, constraints, completed steps, and remaining work. Clients using the same local TOMC database can recall those notes across sessions.
The plugin processes the history supplied to it. It does not automatically collect other chats or remove the current assistant's conversation history. Input savings depend on the task, history, and budget.
Research preview
The current preview supports context preparation through a Python API and assistant integrations, including Codex and Claude Code. The TOMC method has been evaluated on long conversational histories in BEAM and targeted RULER diagnostics, using study-specific evaluation implementations.
TOMC Memory is available on GitHub. See the repository for current installation instructions and assistant integrations. Feedback, use cases, and evaluation collaborations are welcome.
The current source is available under the MIT License.
Related work & software
TOMC Memory is part of a broader research programme on representations. See symbolic representation and sequence compression, De-Time for time-series decomposition, and EchoTime for explainable time-series similarity.