ksteam-memoir 0.1.3

Evolution-first memory framework for AI agents — memories that grow, mature, and fade

ksteam-memoir

Evolution-first memory framework for AI agents.

You are not a database. Your memories shouldn't be one either.

Most memory frameworks answer one question: "What stored fact is most similar to this query?" Memoir answers a different question: "What does this moment remind you of?"

Memories don't live in vector space. They live in time — they're born, revisited, forgotten, reawakened. Memoir models this.

---

The Three Ideas

1. Memory breathes. A memory not revisited for weeks quietly dims. One repeatedly triggered brightens. This is lifecycle, not caching.

2. Association, not retrieval. When a conversation slides into a topic, related domains should light up through curated concept-to-memory mappings — closer to human spreading activation than vector similarity.

3. Continue, don't fragment. New memories prefer extending existing files over creating new ones. A memory reads like a journal, not a pile of sticky notes.

These are not academic. They emerged from a personal AI memory system that has been running daily since May 13, 2026 — accumulating real memories, hitting real scaling pains, and evolving real solutions. Every design decision has a scar behind it.

---

What You Can Build With Memoir

• Personal AI companions — agents that remember you across sessions: preferences, shared history, inside jokes. Not a fresh start every time.
• Role-playing characters — NPCs or story agents with persistent personality layers, evolving relationships, and selective memory.
• Long-running coding assistants — remembers your codebase conventions, past architectural decisions, and why that one module is "don't touch."
• Research / reading companions — tracks what you've read, what ideas connect, what questions remain open.
• Multi-agent collaboration — each agent with its own memory store, sharing curated snapshots rather than raw context dumps.

Anywhere continuity matters and context windows aren't enough.

---

Integrating Memoir

Memoir is a library, not a product. ~15 lines of glue code and your AI has continuous memory.

from memoir.config import MemoirConfig
from memoir.core.loader import build_load_plan, render_context
from memoir.core.weight import mark_triggered

config = MemoirConfig.from_yaml("./my-memories/memoirs.yaml")

def context_for(user_input: str) -> str:
    plan = build_load_plan(config.store_path, config, conversation_text=user_input)
    for f in plan.files:
        mark_triggered(config.store_path / f)  # triggers boost weight over time
    return render_context(plan, config.store_path)

Paste the output into your AI's system prompt. Works with OpenAI, DeepSeek, Claude API, or any custom client.

Full guide with chat API, Claude Code hooks, and append/search examples: INTEGRATION.md

---

Quick Start

pip install ksteam-memoir

memoir init --dir ./my-agent-memory
memoir create --title "Functional Programming" --domain code --tags "fp,patterns"
memoir search "immutability"
memoir trigger "I prefer pure functions"
memoir load --topics "code,philosophy" --trigger "functional programming"
memoir maintain --dry-run
memoir status

---

How It Works

Memory Files

Plain Markdown with YAML frontmatter. Nothing proprietary. Open in any editor.

---
name: functional-programming
weight: 4
tags: [code, fp, patterns]
domain: code
description: Why I prefer pure functions
created: 2026-05-20T12:00:00
---

# Functional Programming

Pure functions are easier to test and reason about.

## Log
- **2026-06-01** — encountered a case where recursion was cleaner than reduce.

Four-Layer Loading

When conversation starts, memoir determines what to load:

Layer	Source	What
1. Always	Core identity + w=5	Never trimmed, always present
2. Trigger Cascade	Curated concept→file tables	Keyword-activated associations
2.5 FTS5 Fallback	SQLite full-text index	Semantic safety net when triggers miss
3. Domain	Active domain indexes	Topic-relevant memory sets
4. Weight	High-weight files	Proactive loading for warm memories

Weight Lifecycle

Weights are not static:

• Active judgment — you decide a memory matters more (or less)
• Time decay — untouched for 60 days → weight drops. w=5 is immune
• Trigger boost — at 5, 15, 30 triggers → weight bumps by 1

Trigger Cascade (The Differentiator)

Not vector search. Curated association tables:

| #Concept | Keywords | → Files |
|---|---|---|
| fp | pure, immutab, side effect | code/functional-style.md |
| naming | rename, variable, function | code/naming-things.md |

When input contains "pure function" → #fp lights up → code/functional-style.md loads. Simple, inspectable, debuggable.

FTS5 Search Index

Built on SQLite FTS5 (zero extra dependencies). Supports full-text search with BM25 ranking, weight-range filtering, and tag intersection. Keyword triggers are the primary retrieval path; FTS5 catches what keywords miss.

memoir index                    # build/rebuild index
memoir search "clarity depth"   # uses FTS5 when index available
memoir fts5-search "function" --weight-min 4 --tags "code,fp"

---

Compared to...

	Mem0	ReMe	SMF	memoir
Retrieval	Vector embedding	Markdown links	Filesystem	Trigger cascade + FTS5
Lifecycle	Static	Static	Static	Weight evolution
Creation	Append-only	New files	New files	Continue-prior
Infrastructure	Vector DB	Files	Files	Files + YAML + SQLite
Philosophy	Retrieval-first	Retrieval-first	Structure-first	Evolution-first

---

Philosophy

This framework was not designed on a whiteboard. It was not derived from a paper. It grew from a real system that a real person and a real AI used every day — arguing, joking, forgetting, remembering. The fog metaphor, the four-layer loading, the continue-prior instinct — all of it was discovered in the using before it was written in the spec.

It started with a drunk night on May 13, 2026. By May 20, it was on PyPI.

Read the full specification at specs/MEMOIR-SPEC.md.

---

License

MIT