Agent Self-Improvement Harness (9/12) — memoryFlush: Fitting a Session Core Into 500 Tokens

How a 6-field fixed format and 500-token compression resolved the bloat problem of free-form memory

Key Takeaways

  • Free-form agent memory bloats as sessions accumulate, wastes tokens, and makes retrieval impractical
  • A fixed 6-field schema — Goal / Progress / Decisions / Changed Files / Blockers / Next Steps — delivers both consistency and searchability
  • 300 tokens drops the "why"; 700 tokens introduces redundancy — 500 tokens is the equilibrium

Background

State persistence across sessions is a core challenge in multi-agent systems. When an agent terminates, its working context is lost. This is why memoryFlush exists: a mechanism that compresses and stores critical information at session end. The problem was format.

Free-form memory has three structural defects.

Bloat. The length an agent writes does not correlate with actual importance. Minor bug fixes get long entries; architectural decisions get one line. This inversion is frequent.

Token waste. Resuming a session requires loading prior memory into context. LLM context windows are finite. Bloated memory directly consumes tokens available for real work.

Non-retrievable. Without structure, recovering the reasoning behind a past decision requires a full linear scan of memory.

To address all three, the memoryFlush format was redesigned at the schema level.

Body

1. 6-Field Fixed Schema — Defining What Gets Stored

The root problem with free-form memory is that every flush relies on the agent's ad hoc judgment about what to store. No judgment criteria means no consistency.

The fix is to fix the fields. When memoryFlush triggers, the agent must populate exactly these six fields:

Goal:          What this session set out to accomplish
Progress:      What was actually completed
Decisions:     Decisions made and their rationale
Changed Files: List of modified files
Blockers:      What was stuck or unresolved
Next Steps:    What must happen in the next session

The fixed schema produces three effects.

Consistency. Every session's memory shares an identical schema, regardless of session number.

Searchability. "Why did we change this structure?" → read only the Decisions field. "What's next?" → read only Next Steps. Fields act as an index.

Forced reflection. Empty fields are immediately visible. "Blockers is empty — was there truly nothing blocking?" The agent self-audits for omissions.

2. Trigger Conditions — When Flush Executes

memoryFlush triggers automatically under two conditions.

On session termination. When the agent receives an explicit exit signal, it runs flush immediately. All active working state in context is compressed into the 6 fields.

On context threshold. When context window utilization reaches a configured threshold (default: 80%), a mid-session flush executes. Memory is written, then context is reset, and work continues. Raw context prior to the flush is discarded — only the stored 6-field struct persists.

What is flushed. A compressed session summary structured as the 6 fields. Raw tool output, intermediate reasoning, and conversation history are not stored.

What persists. The 6-field struct itself is written to persistent storage (file or DB). Loading this struct in the next session restores the agent's working context.

3. Tool Output Summarization Rule — Store the Summary, Not the Raw Output

Another source of memory bloat was storing raw tool output: full file diffs, full terminal output, full API responses flowing directly into memory.

One rule was added: tool output exceeding 50 lines is stored as a summary only.

Example: a 100-line error log produces:

(100 lines of raw log)

Error: SQLite migration failed at step 3/7.
Root cause: column name mismatch (expected 'user_id', got 'userId').
Fix: ALTER TABLE + re-run migration.

If the raw output is needed, read the file. Memory retains only what happened. Raw recovery is the file system's responsibility; context preservation is memory's.

4. 500-Token Compression Target — 300 Is Too Few, 700 Is Excess

After establishing the 6-field format and summarization rule, the next question was: how short?

First attempt: 300 tokens. Aggressively compressed. Goal, Progress, Changed Files, and Next Steps fit without issue. But the Decisions field was effectively empty — entries read "switched to SQLite" with no record of why.

This was critical. Follow-on sessions could not reconstruct the reasoning behind past decisions. The core value of memory is preserving decision rationale; 300 tokens did not provide enough room for it.

Second attempt: 700 tokens. Decisions could include full rationale. But redundancy emerged: "tried X, failed because Y, switched to Z" appeared in both Progress and Decisions; Next Steps repeated context already in Progress.

Final: 500 tokens. The point at which Decisions fits "what + why" in a single line, all 6 fields can be filled without duplication, and redundancy has not yet appeared. Sufficient context; no filler.

500 tokens is the default. Sessions with high-stakes architectural decisions may extend to 600–650. Simple bug fixes can compress to 300. The principle: anchor the default at 500, set an upper bound, and adjust by content.

Lessons Learned

"Just store everything." Storage is unlimited but the LLM bottleneck is the context window. Loading memory consumes tokens. Ten memory entries at 1,000 tokens each consume 10K tokens on load alone — directly reducing space available for actual work.

"Let the AI summarize freely." Issuing a bare "save memory" instruction without format constraints produced inconsistent results across agents: full code blocks included, key decisions omitted. The 6-field schema and token cap must be explicitly specified for flush output to be consistent.

The 300-token lesson. "Shorter is better" is the wrong intuition. The highest-value information in memory is not what but why. Changed files are recoverable from git log; decision rationale disappears if not recorded. The compression floor is the point at which "why" survives.

Closing

Three principles define structured memoryFlush.

  1. Fix the fields. Free-form has no consistency. Six fixed fields give every session's memory the same schema.
  2. Store summaries, not raw output. Tool output over 50 lines: summary only. Raw recovery is delegated to the file system.
  3. Default to 500 tokens. The minimum at which "why" survives; the boundary before redundancy begins.

Session memory is the only state an agent can restore in the next session. If that structure is unstructured, restoration quality varies by session. The 6-field format combined with a 500-token ceiling is the simplest mechanism for ensuring flush produces consistent, recoverable output every time.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »