Problem
Maka currently provides a local MEMORY.md / PENDING.md workflow for storing and injecting approved memory entries.
This provides basic persistent memory, but it still requires explicit maintenance and does not form a complete long-term memory lifecycle. In particular, Maka cannot yet automatically:
- extract durable information from completed sessions;
- consolidate related information across sessions;
- recall more detailed memory when
MEMORY.md is insufficient;
- collect feedback about whether recalled memory was useful or incorrect;
- correct or forget outdated memory at a sufficiently fine granularity.
Treating an entire session as the smallest memory unit would also be too coarse. A session may contain several independent memory points, and invalidating one point should not remove everything learned from that session.
Desired outcome
Maka should have an automatic background memory pipeline that:
- extracts durable information from eligible sessions;
- stores independently manageable memory points;
- maintains
MEMORY.md and memory_summary.md;
- supports keyword-based recall;
- records recall feedback;
- corrects or forgets individual memory points without deleting unrelated memories;
- keeps existing
RuntimeEvent data as the original evidence source.
The first implementation should focus on the memory lifecycle itself. Embedding-based or vector retrieval can be considered separately.
Alternatives or workarounds
The existing MEMORY.md / PENDING.md workflow can continue to support manually reviewed memory entries, but it requires explicit user maintenance and does not provide automatic extraction, consolidation, feedback, or forgetting.
中文翻译
问题
Maka 当前提供了基于 MEMORY.md 和 PENDING.md 的本地记忆流程，用于保存并注入经过批准的记忆。
这已经提供了基础的持久化记忆能力，但仍然需要显式维护，尚未形成完整的长期记忆生命周期。Maka 目前还不能自动完成：
- 从已结束的 Session 中提取具有长期价值的信息；
- 跨 Session 整理和合并相关信息；
- 当
MEMORY.md 信息不足时读取更详细的记忆；
- 记录召回的记忆是否有用或存在错误；
- 以足够细的粒度修正或遗忘过时记忆。
如果把整个 Session 作为最小记忆单位，粒度也会过粗。一个 Session 可能包含多个独立记忆点，其中一条失效时，不应该删除这个 Session 中的所有记忆。
期望结果
Maka 应当具有自动运行的后台记忆流程，能够：
- 从符合条件的 Session 中提取长期信息；
- 保存可以独立管理的记忆点；
- 维护
MEMORY.md 和 memory_summary.md；
- 支持关键词召回；
- 记录记忆使用反馈；
- 独立修正或遗忘某个记忆点；
- 保留现有
RuntimeEvent 作为原始证据来源。
第一版应优先实现完整的记忆生命周期。Embedding 或向量检索可以作为独立功能后续考虑。
替代方案或临时做法
当前 MEMORY.md / PENDING.md 流程仍然可以支持人工审核的记忆，但需要用户显式维护，也不能自动完成提取、整理、反馈和遗忘。
Problem
Maka currently provides a local
MEMORY.md/PENDING.mdworkflow for storing and injecting approved memory entries.This provides basic persistent memory, but it still requires explicit maintenance and does not form a complete long-term memory lifecycle. In particular, Maka cannot yet automatically:
MEMORY.mdis insufficient;Treating an entire session as the smallest memory unit would also be too coarse. A session may contain several independent memory points, and invalidating one point should not remove everything learned from that session.
Desired outcome
Maka should have an automatic background memory pipeline that:
MEMORY.mdandmemory_summary.md;RuntimeEventdata as the original evidence source.The first implementation should focus on the memory lifecycle itself. Embedding-based or vector retrieval can be considered separately.
Alternatives or workarounds
The existing
MEMORY.md/PENDING.mdworkflow can continue to support manually reviewed memory entries, but it requires explicit user maintenance and does not provide automatic extraction, consolidation, feedback, or forgetting.中文翻译
问题
Maka 当前提供了基于
MEMORY.md和PENDING.md的本地记忆流程，用于保存并注入经过批准的记忆。这已经提供了基础的持久化记忆能力，但仍然需要显式维护，尚未形成完整的长期记忆生命周期。Maka 目前还不能自动完成：
MEMORY.md信息不足时读取更详细的记忆；如果把整个 Session 作为最小记忆单位，粒度也会过粗。一个 Session 可能包含多个独立记忆点，其中一条失效时，不应该删除这个 Session 中的所有记忆。
期望结果
Maka 应当具有自动运行的后台记忆流程，能够：
MEMORY.md和memory_summary.md；RuntimeEvent作为原始证据来源。第一版应优先实现完整的记忆生命周期。Embedding 或向量检索可以作为独立功能后续考虑。
替代方案或临时做法
当前
MEMORY.md/PENDING.md流程仍然可以支持人工审核的记忆，但需要用户显式维护，也不能自动完成提取、整理、反馈和遗忘。