"Memory Systems — Preserving Information Outside the Context Window (Harness Series 3/6)"

Part 2 argued for small context windows. So where does the older information go? Into external memory systems. By 2026, memory itself has become its own infrastructure category.

This article covers agent memory — short-term vs long-term, three types, vector vs graph, solutions like Mem0 and Zep, and where SQLite+FTS5 ends and Pinecone begins.

Series Roadmap (6 parts)

  1. What Is Harness Engineering?
  2. Context Engineering
  3. Memory Systems ← this article
  4. Tools & Sandboxing
  5. Multi-Provider Routing
  6. Evaluation & Ops

1. Memory ≠ Extended Context Window

A common confusion: "context is memory."

The accurate split: - Context = the model input on the current turn (volatile, costs tokens) - Memory = external storage that persists across sessions (durable, queryable)

Context fills fresh every turn. Memory is queried every turn — only relevant fragments get injected into context.

2026 consensus: "Memory is a first-class architectural component." It now has dedicated benchmarks (LoCoMo, LongMemEval), its own research literature, measurable performance gaps, and a tool ecosystem.

2. Three Types of Memory

2-1. Episodic

What happened. Time-ordered events. - "User created the X function yesterday" - "Last session refactored the auth module"

2-2. Semantic

What is known. Time-invariant facts. - "User is a Python developer" - "Project uses PostgreSQL"

2-3. Procedural — added in 2026

How to do things. Repeatable procedures. - "PR procedure for this project: lint → test → review → merge" - "This user wants conventional-commit format checked before commit"

Mem0's v1.0.0 API explicitly added this type in 2026. Earlier APIs covered only episodic + semantic.

3. Vector vs Graph vs SQL/FTS5

Vector Memory (semantic search)

  • Embed the query → retrieve top-N most similar memories
  • Strength: semantic proximity, fuzzy match
  • Weakness: weak at exact relationships

Example: "User mentioned Python" — retrieves relevant memories well.

Graph Memory (relational reasoning)

  • Stores events, entities, and relationships as a graph
  • Strength: multi-hop reasoning ("X is part of Y, which is at Z")
  • Weakness: build/maintenance cost ↑

Example: "User is a Python developer building data pipelines with pandas, at a company that uses dbt and is migrating off Spark" — connected facts retrievable.

SQL + FTS5 (keyword + structured)

  • SQLite's FTS5 module is a fast, lightweight keyword search
  • Strength: zero infrastructure (single file), 4,300-memory query in <1ms
  • Weakness: weak fuzzy match, no embedding semantics

Pinecone p95 latency: 25~50ms. SQLite+FTS5 at the same scale: under 1ms. Pure keyword wins on SQL.

Hybrid (RRF — Reciprocal Rank Fusion)

  • Vector top-N + FTS5 top-N → fused via RRF
  • Best of both
  • OpenClaw memcore and wikycore implement this pattern

4. Solution Comparison (April 2026)

Mem0 (managed)

  • Position: User personalization + cost efficiency
  • Core tech: Memory Compression Engine — compresses chat history into dense representations
  • Effect: claims 80% prompt-token reduction
  • Pricing: Managed SaaS. Self-host also possible (open source).

Zep (graph)

  • Position: Time-aware knowledge graph
  • Core tech: Temporal Knowledge Graph — stores time validity of facts
  • Strength: Time-dependent queries like "User used to use Python, now uses Rust"
  • Latency: Sub-second

Memanto (academic)

  • arXiv 2604.22085
  • "Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents"
  • Information-theoretic retrieval — ranks by uncertainty reduction

MemMachine (academic)

  • arXiv 2604.04853
  • "Ground-Truth-Preserving Memory System for Personalized AI Agents"
  • Compression policy that prioritizes preservation of original information

memcore (OpenClaw)

  • SQLite + vector + RRF + 3-tier memory layers
  • Plan / Lessons / Sessions pattern
  • v0.5 introduced MemPalace techniques (location-based recall)

5. Production Patterns

5-1. 3-Layer Memory (CLAUDE.md standard)

Layer 1: Project Knowledge (docs/) — permanent, keyword-indexed
Layer 2: Behavioral Rules (lessons.md) — what to do / not do
Layer 3: Session Restore (sessions/) — volatile snapshots

CLAUDE.md recommends this 3-layer approach. Claude Code, OpenClaw, and Hermes all use this skeleton.

5-2. Auto-Harvesting

End-of-session automatically stores new learnings to the right layer: - Error resolution → lessons.md - New domain knowledge → docs/{category}/ - Progress state → sessions/

CLAUDE.md's "Automatic Memory Harvesting (Instinct Pattern)" section formalizes this.

5-3. Dedup + Cascade Update

  • Same pattern repeated 2× → promote to rule (increment lessons.md count)
  • New info conflicts with existing memory → flag conflict + ask user
  • One fact changes → check memory-map.md for connected files

5-4. Eviction Policy

Unbounded accumulation = noise. Example policy: - Time-based: archive after 90 days unused - Use-based: remove if count < 2 and unused > 30 days - LRU: under context pressure, drop oldest first

6. SQL vs Vector — Where to Start?

Starting (personal project, 1 user)

SQL + FTS5 only - Single SQLite file - 100~10,000-memory range is plenty - Zero infrastructure, backup = file copy - OpenClaw memcore's early version started here

Mid-scale (multi-user, 10K+ memories)

SQL + FTS5 + vector (hybrid) - Add Chroma or Qdrant - bge-m3 embeddings (free locally on oMLX) - Combine results via RRF

Large (millions of memories)

Mem0 / Zep / custom infrastructure - Mem0: automated user-persona extraction - Zep: temporal graph + complex relationships - Custom: ElastiCache + Neptune Analytics on AWS

7. Anti-Patterns

7-1. "Skipping vector DB is amateur"

At 4,300 memories: SQLite+FTS5 = <1ms, Pinecone p95 = 25~50ms. SQLite has zero infrastructure cost. Start with SQL.

7-2. "Just dump everything; we can search later"

Unbounded accumulation → noise. Without an eviction policy, retrieval accuracy decays after a year.

7-3. "Write memory once and forget it"

Memory is a living system. Without update / conflict resolution / archive flows, stale memory leads to wrong answers.

CLAUDE.md: "memory records can become stale over time. Verify before using."

7-4. "Pick an embedding model and never change it"

Switching embedding models invalidates all existing vectors. Migration cost is huge. Pick a proven one (bge-m3, voyage-3) up front.

8. Knowledge Graphs in Detail

Graph memory is rising in 2026 because: - LLMs are weak at multi-hop reasoning (vector retrieval limit) - Relations between facts don't reduce to similarity - Time dimension — when a fact was true

KG Memory Build Flow

  1. Input (conversation/document) → entity extraction
  2. Relations between entities → LLM-driven or NER + relation classifier
  3. Store in graph (Neo4j, Neptune, custom SQLite)
  4. Query starts from a node → graph traversal → connected facts

Cost vs Value

  • Build: LLM call per input → cost ↑
  • Search: traversal is fast
  • Value: qualitative gain on multi-hop queries

Recommendation: For per-user personal assistants, KG memory (Zep) is worth it. For single-task agents (e.g., a coding agent), hybrid (SQL+vector) is sufficient.

Bottom Line

Memory decision Recommendation
Starting point SQLite + FTS5
Add semantic search + vector (Chroma/Qdrant) + bge-m3
Need relational reasoning KG (Zep)
User personalization Mem0
Academic frontier Memanto / MemMachine

The single takeaway: "Memory is not an extension of context — it is a separate system."

Part 4 (next) covers what the agent actually does — Tools & Sandboxing.

First-Party Sources

  • Mem0 State of AI Agent Memory 2026: mem0.ai/blog/state-of-ai-agent-memory-2026
  • Memanto: arxiv.org/abs/2604.22085
  • MemMachine: arxiv.org/abs/2604.04853
  • AWS Mem0 + Neptune: aws.amazon.com/blogs/database/build-persistent-memory-for-agentic-ai
  • Zep Temporal Knowledge Graph: getzep.com

๋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 »