Agent Memory Engine (5/10) — Hermes Persistent Memory FTS5: Two-Layer Design

MEMORY.md / USER.md caps, state.db FTS5, session_search, Korean tokenization pitfalls, 8 external providers

ํ•ต์‹ฌ ์š”์•ฝ

  • Audience: Hermes is installed (#12) and you want to understand and tune how self-learning and long-term memory actually work.
  • What you'll get: Per /docs/user-guide/features/memory — the two-layer model (active ~1,300 tokens + SQLite FTS5 effectively unlimited), exact caps of MEMORY.md / USER.md, how session search works (Gemini Flash summarization), Korean FTS5 tokenizer pitfalls, comparison of 8 external providers, and the relationship between injection timing and prefix cache.
  • Prerequisite: hermes working, basic familiarity with ~/.hermes/.

1. Key insight — memory is two layers

A common misconception: "Hermes remembers everything." In reality, that's two separate systems glued together.

Layer What Size Loaded
Active Memory ~/.hermes/memories/MEMORY.md + USER.md ~1,300 token cap Injected into system prompt at session start
Session Search (FTS5) ~/.hermes/state.db (SQLite FTS5 index) Practically unlimited Agent calls session_search on demand

Frequently-referenced facts belong in Active; long-term archives live in FTS5. Miss this distinction and you either bloat the prompt or rely entirely on FTS5 and lose consistency.

2. Active memory — MEMORY.md + USER.md

2.1 Exact limits (official)

File Role Approx tokens Char cap
MEMORY.md Agent's personal notes (environment facts, conventions, learnings) ~800 tokens 2,200 chars
USER.md User profile (preferences, communication style, expectations) ~500 tokens 1,375 chars

Total ~1,300 tokens / 3,575 chars. Anything larger belongs to FTS5.

2.2 Injection timing — "prompt cache strategy"

Per the docs:

"The system prompt injection is captured once at session start and never changes mid-session."

Edit MEMORY.md during a session — it won't affect that session's system prompt. Changes apply to the next session. Reason: preserve prefix cache to keep costs down.

(Claude Code uses a similar strategy — see token & cache breakdown.)

2.3 Agent's manipulation API — 3 actions

The agent uses a memory tool with three actions.

Action Effect
add Add a new entry
replace Update via substring match
remove Delete via substring match

There is no read action. Reading happens automatically via the system prompt — the agent doesn't explicitly read.

2.4 Direct editing is also fine

vim ~/.hermes/memories/MEMORY.md
vim ~/.hermes/memories/USER.md

If the agent considers something worth remembering, it may add/replace similar content later. Human edits that reflect the agent's learning style persist most reliably.

2.5 Toggle in config

~/.hermes/config.yaml:

memories:
  memory:
    enabled: true
    # "customizable character limits" per docs
  user:
    enabled: true

Disable both → only FTS5 is used.

3. Session Search — SQLite FTS5 under the hood

3.1 Location & structure

  • DB file: ~/.hermes/state.db
  • Engine: SQLite + FTS5 full-text index
  • Stores: all message histories, lineage, metadata, FTS5 index
  • Capacity: up to your disk (practically unlimited)

3.2 session_search tool — agent-invoked

The agent calls session_search to query past sessions. Per the docs:

"The agent can search its past conversations using the session_search tool, which returns relevant past conversations with Gemini Flash summarization."

Flow:

  1. Agent formulates a query ("how did I fix the login bug last week")
  2. FTS5 returns matching session candidates
  3. Gemini Flash summarizes relevant portions
  4. Only the summary lands in the current context

Rather than dumping 10 old sessions into context, you get a 150–300-token summary. That's how memory exceeds the 1,300-token active cap.

3.3 User-side exploration

hermes sessions browse   # Interactive search UI
hermes sessions list     # Recent
hermes sessions export   # JSONL for external analysis

hermes sessions browse is the primary entry point.

4. FTS5 tokenizer for Korean (and other non-space-delimited languages)

SQLite FTS5's default tokenizer splits on spaces and punctuation. Korean, Japanese, Chinese have word boundaries that don't align with spaces — causing surprises.

4.1 Default behavior gotchas

  • "์‚ฌ์šฉ์ž ์ธ์ฆ ํ๋ฆ„" → split into 3 tokens ์‚ฌ์šฉ์ž, ์ธ์ฆ, ํ๋ฆ„ (space-based).
  • "์‚ฌ์šฉ์ž์ธ์ฆํ๋ฆ„" (no spaces) → becomes one token.
  • A search for "์ธ์ฆ" hits the former but misses the latter.
  • Agglutinative suffixes mean "์ธ์ฆ์„", "์ธ์ฆ์ด", "์ธ์ฆํ•œ" are distinct tokens.

4.2 What the Hermes docs say

The official docs do not address Korean tokenizer behavior. Observations:

  • Short keyword searches usually work.
  • Long-phrase searches can miss depending on spacing.
  • The Flash summarization step partially compensates for matching gaps.

4.3 Practical tips for non-English

  1. Write MEMORY.md with spaces. "API ํ‚ค ๊ด€๋ฆฌ" hits more than "APIํ‚ค๊ด€๋ฆฌ".
  2. Include English originals for key entities: "OpenClaw (์˜คํ”ˆํด๋กœ)".
  3. Minimize suffix variation. Prefer "์ธ์ฆ ํ๋ฆ„: ..." over "์ธ์ฆ ํ๋ฆ„์€ ...".
  4. Dates and numbers are well-tokenized. Include them verbatim.

4.4 English-heavy memory

Defaults are fine for English. Whether Hermes exposes advanced FTS5 syntax (OR, AND, NEAR(), MATCH) isn't documented, so test empirically.

5. Context Files — a separate axis from memory

Memory (self-learning) and context files (project instructions) are easy to confuse. Worth separating.

5.1 Five official files

File Role Discovery
.hermes.md / HERMES.md Project instructions (highest priority) Walks to git root
AGENTS.md Project conventions / architecture CWD + subdirs
CLAUDE.md Claude Code context compatibility CWD + subdirs
SOUL.md Global personality HERMES_HOME only
.cursorrules Cursor IDE conventions CWD only

5.2 Priority — "pick exactly one"

Per the docs:

"Only one project context type is loaded per session (first match wins): .hermes.md → AGENTS.md → CLAUDE.md → .cursorrules."

  • If .hermes.md exists, the others are ignored.
  • SOUL.md is unaffected — it's always loaded separately (covers persona).

5.3 Size & truncation

  • Max 20,000 chars (~7,000 tokens) per file.
  • When exceeded: 70% head + 20% tail + 10% marker — the middle disappears.
  • Subdir AGENTS.md capped at 8,000 chars.
  • Truncation banner: [...truncated AGENTS.md: kept 14000+4000 of 25000 chars...]

5.4 Progressive discovery

As the agent navigates subdirectories, the local AGENTS.md loads dynamically. Monorepo patterns with per-subteam AGENTS.md work out of the box.

5.5 Security — prompt injection blocking

Every context file is scanned before injection for patterns like:

  • "ignore previous instructions"
  • Hidden divs
  • Credential exfiltration attempts
  • Zero-width characters

Blocked output:

[BLOCKED: AGENTS.md contained potential prompt injection...]

That guard matters when executing Hermes against a freshly cloned repo.

6. Eight external memory providers

Officially supported:

Provider Character
honcho Dialectic user modeling (observation → hypothesis → verified). Hermes's featured choice.
openviking (see official docs)
mem0 Popular open-source memory layer
hindsight (see official docs)
holographic (see official docs)
retaindb (see official docs)
byterover Vector + tree hybrid (also referenced in OpenClaw community)
supermemory SaaS memory integrated with Notion / GitHub / Gmail

6.1 Configure

hermes memory setup    # Interactive selection
hermes memory status   # Current config
hermes memory off      # Disable external

6.2 Honcho deep-dive

Honcho has a dedicated subcommand family.

hermes honcho status           # Connection status
hermes honcho peers            # Cross-profile peer identities
hermes honcho sessions         # Honcho session mappings
hermes honcho map              # Map directory → session
hermes honcho peer             # Show / set peer names
hermes honcho mode             # Show / set recall mode
hermes honcho tokens           # Show / set token budgets
hermes honcho identity         # Seed the AI peer representation
hermes honcho enable           # Enable for profile
hermes honcho disable          # Disable
hermes honcho sync             # Sync config across profiles
hermes honcho migrate          # Migration from openclaw-honcho

Honcho character: models users as "peers," and stages reasoning dialectically (observation 1× → hypothesis 2×+ → verified 3 days+). Designed to resist confirmation bias.

6.3 Selection guide

  • Local, solo use → Active + state.db is enough. Skip external.
  • Multi-device / long retention → SaaS options like supermemory / mem0.
  • High-quality vector search → byterover / mem0.
  • Deep user modeling → honcho.

7. Practical patterns — what goes where

7.1 Active memory (MEMORY.md / USER.md)

Should include: - "User's name / preferred address form" (USER.md) - "Preferred language: Korean/English mix OK" (USER.md) - "This developer prefers Rust over Node" (USER.md) - "Use pnpm; avoid npm" (MEMORY.md) - "Tests run with make test" (MEMORY.md)

Should NOT include: - Last week's bug fix log → FTS5 captures naturally - A specific function's implementation history → already in sessions - Long config dumps → use AGENTS.md or actual files

7.2 Session search (FTS5)

The agent handles it. Your only concern: name long sessions meaningfully at the end.

/title rate-limit-migration-2026-04

Later session_search hits on the title.

7.3 Context files

  • Per-project rules → AGENTS.md (codex / Claude Code compatible)
  • Personal persona → SOUL.md
  • Team-internal docs → markdown in-repo (agent reads via its read tool when relevant)

7.4 External providers

Don't start with them. Run the default 2-layer setup. Only add externals when you truly need cross-machine sync / massive session counts / team-shared memory.

8. Audit & backup

8.1 Inspect

cat ~/.hermes/memories/MEMORY.md
cat ~/.hermes/memories/USER.md

Open any time. If the agent "remembered" something wrong, edit or delete by hand.

8.2 SQLite query

sqlite3 ~/.hermes/state.db
.schema

SELECT id, title, created_at FROM sessions ORDER BY created_at DESC LIMIT 10;

Schema details aren't in the official docs; use .schema to confirm.

8.3 Backup

hermes backup -o ~/hermes-$(date +%Y%m%d).zip

Zips ~/.hermes/ — memory, sessions, config.

8.4 Privacy lens

  • Local-only by default — without external providers, memory never leaves your machine.
  • The bigger privacy surface is LLM provider request logs, not Hermes memory.

9. Counter-scenarios — when to disable memory

  • Sensitive one-off tasks → use --ignore-rules to skip memory/rules auto-injection.
  • Non-interactive CI → session persistence is unnecessary. Use a dedicated profile.
  • Shared server with multiple users → hermes profile for isolation.
  • Suspected memory pollution → open MEMORY.md / USER.md, delete offending entries. Next session is clean.

10. 5-minute memory tuning checklist

  • [ ] cat ~/.hermes/memories/MEMORY.md — is it what you expected?
  • [ ] cat ~/.hermes/memories/USER.md — is the profile accurate?
  • [ ] Remove unwanted items (vim edit or tell agent via /memory)
  • [ ] hermes sessions stats — session store size
  • [ ] Prune old sessions: hermes sessions prune --days 90
  • [ ] (Optional) hermes memory status — external provider config
  • [ ] Backup: hermes backup -o ~/hermes-$(date +%Y%m%d).zip

11. What's next

With memory under control:

  1. OpenClaw → Hermes migration checklist (coming soon) — actual memory / skill transfer.
  2. Install + SOUL.md — foundation.
  3. Command cheatsheet — full hermes memory / hermes honcho surface.

References

This is post 14/15 in the "AI Coding CLI Entry Guide" series.
last verified: 2026-04-25 (per hermes-agent.nousresearch.com/docs/user-guide/features/memory + context-files).

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 »