Ledger Memory System
A cross-session experience accumulation mechanism for AI agents that consolidates error analyses, retrospective lessons, and design decisions from each conversation into reusable behavioral rules. It addresses the lack of cross-session experience accumulation found in mainstream agent framework Session, Context, and Memory mechanisms.
Timeline (last 90 days)
The Ledger memory system aims to address the pain point that off-the-shelf Session/Context/Memory mechanisms lack the ability to accumulate experience across sessions.
The author encapsulated the Ledger into a Skill and ported it to DeepSeek Harness, allowing users to trigger viewing it by saying 'check the ledger.'
The author's Ledger accumulated 24 iron rules, 2 preferences, 18 drafts, 4 solidified trajectories, and 14 collected pieces of information over two to three days of use.
The 'iron rules' in the Ledger are not manually hard-coded but instead gradually evolve and naturally accumulate from errors and retrospectives during use.
The concept behind the Ledger system shares similarities with the WikiSkill (evolutionary system) proposed by Google researchers, but the Ledger was independently designed and developed by the author.
The underlying layer of the Ledger system is a pure document storage layer, where all iron rules, preferences, drafts, and collected information are essentially document-type storage.
The two pathways for data to enter the Ledger are: automatic capture of tool errors via a Hook mechanism, and manual retrospectives initiated by user commands.
The cross-framework portability of the Ledger depends on the Agent framework providing Hook capabilities, with the remaining functionality implemented through contracts and tools, decoupled from the specific orchestration engine.
The contract mechanism of the Ledger system requires the Agent to generate a summary after completing a task and place it in the observation area of the Ledger.
Due to the limited context window, the Ledger cannot inject all records at once, so the author adopts a 'drawer-style' layered injection scheme that only pulls relevant content layer by layer as needed.
2 more timeline events
All Facts (11)
The Ledger memory system aims to address the pain point that off-the-shelf Session/Context/Memory mechanisms lack the ability to accumulate experience across sessions.
50%UnverifiedThe author encapsulated the Ledger into a Skill and ported it to DeepSeek Harness, allowing users to trigger viewing it by saying 'check the ledger.'
50%UnverifiedThe 'iron rules' in the Ledger are not manually hard-coded but instead gradually evolve and naturally accumulate from errors and retrospectives during use.
50%UnverifiedThe concept behind the Ledger system shares similarities with the WikiSkill (evolutionary system) proposed by Google researchers, but the Ledger was independently designed and developed by the author.
50%UnverifiedThe underlying layer of the Ledger system is a pure document storage layer, where all iron rules, preferences, drafts, and collected information are essentially document-type storage.
50%UnverifiedThe two pathways for data to enter the Ledger are: automatic capture of tool errors via a Hook mechanism, and manual retrospectives initiated by user commands.
50%UnverifiedThe cross-framework portability of the Ledger depends on the Agent framework providing Hook capabilities, with the remaining functionality implemented through contracts and tools, decoupled from the specific orchestration engine.
50%UnverifiedThe contract mechanism of the Ledger system requires the Agent to generate a summary after completing a task and place it in the observation area of the Ledger.
50%UnverifiedDue to the limited context window, the Ledger cannot inject all records at once, so the author adopts a 'drawer-style' layered injection scheme that only pulls relevant content layer by layer as needed.
50%UnverifiedThe author recommends that after obtaining the Ledger Skill (whether using Claude Code or Codex), one should not register it directly as a Skill but should first have the model analyze and tune it before registering and using it.
50%UnverifiedThe author believes that off-the-shelf Agent frameworks lack sufficient 'mindset,' missing standardized processes and actions, which causes the Ledger Skill to underperform after porting.
50%