MADE:Explorations · Features

MemPalace

The agent that remembers everything.
Five-tier persistent memory with 96.6% recall on LongMemEval.

Active · MIT License
May 2026 · colombod/amplifier-bundle-memory
The Problem

AI agents
have amnesia

🧠

Context Window Limits

Every session starts from zero. Agents can't recall what you discussed yesterday, last week, or last month.

🔄

Repeated Explanations

You re-explain your preferences, project decisions, and architectural choices in every single conversation.

🔍

Lost Discoveries

Gotchas, debugging insights, and hard-won knowledge vanish when the session ends. Future sessions repeat past mistakes.

An agent without memory is a brilliant colleague with permanent anterograde amnesia — every conversation is their first day on the job.

Architecture

Five tiers of memory,
one cohesive system

🔮

Semantic Vectors

Verbatim storage with embedding-based similarity search. 96.6% R@5 retrieval.

🕸️

Knowledge Graph

Subject-predicate-object triples. Cross-wing traversal and relationship queries.

📋

Coordination Files

HANDOFF.md, GLOSSARY.md, PROJECT_CONTEXT.md — structured markdown that survives git clones.

🗄️

SQLite Fact Store

FTS5 full-text search for explicit key-value memories. Fast structured lookups.

📓

Agent Diaries

Per-agent session journals. What was worked on, what was learned, what comes next.

Each tier excels at a different retrieval pattern. The palace metaphor — wings, rooms, drawers — gives agents an intuitive spatial navigation model across all five.

How It Works

Memory flows through
the entire session

1

Session Briefing

hooks-mempalace-briefing assembles a wake-up briefing from semantic search, knowledge graph, agent diary, and HANDOFF.md. Importance re-ranking surfaces what matters most.

2

Auto-Capture

hooks-mempalace-capture files verbatim tool outputs as palace drawers with auto-detected wing and room. No manual "remember this" required.

3

Mid-Session Interject

hooks-mempalace-interject surfaces relevant memories at the right moment. Only fires when cosine similarity ≥ 0.72 — non-disruptive by design.

4

End-of-Session Curation

The Curator agent runs three phases: palace curation, coordination file updates (HANDOFF.md, PROVENANCE.md), and KG enrichment with importance scoring and duplicate linking.

Benchmarks

96.6%

Recall@5 on LongMemEval — raw semantic retrieval, no LLM post-processing.

96.6
R@5 — Raw
LongMemEval
98.4
R@5 — Hybrid v4
Held-out set
88.9
R@10 — LoCoMo
Hybrid v5
Benchmarks are properties of MemPalace's retrieval engine.
Full methodology in docs/research/gene-transfer-v1.2.0.pdf.
Key Innovation

Memory that speaks up
at the right moment

The interject hook surfaces relevant memories mid-session — but only when it's genuinely useful.

Non-Disruptive Design

Fires on prompt_submit, tool_pre, and orchestrator_complete events. Requires cosine similarity ≥ 0.72 before surfacing anything. In the uncertain band, an LLM judge decides relevance.

Event Observability

Every hook emits structured events to ~/.mempalace/events/{session_id}.jsonl. Query with palace events or tail -f for live observability. Kill switch per hook: emit_events: false.

The worst memory system is the one that interrupts you with irrelevant trivia. MemPalace's interject hook stays silent unless it has something genuinely worth saying.

Agents

Three specialized agents,
clear separation of concerns

🔍

Archivist

The read path. Precise palace search, KG queries, graph traversal, and coordination file reads. Invoked on-demand when you need to find something.

🗣️

Docent

Conversational memory Q&A. Ask natural-language questions: "What decisions have I made about auth?" or "Summarize last week's work."

✍️

Curator

The write path. Palace curation, Phase 3 KG enrichment with importance scoring, HANDOFF.md updates, and duplicate linking. Runs at session end and on-demand.

Palace Garden

On-demand deep structural analysis: BFS clustering, KG edges, diary entries, and importance backfill. Zero deletion — all outputs are additive KG facts, capped at 500 drawers with a 120s timeout.

Cross-Platform Files

Coordination files (PROJECT_CONTEXT.md, GLOSSARY.md, HANDOFF.md) are readable by Amplifier, OpenAI Codex, GitHub Copilot, Cursor, and Windsurf. Memory that travels with the repo.

Quick Start

Three steps to total recall

# 1. Install the MemPalace engine pip install mempalace # 2. Add the bundle to Amplifier amplifier bundle add git+https://github.com/colombod/amplifier-bundle-memory@v1.2.1 amplifier bundle use memory # 3. Run — coordination files scaffold automatically amplifier run "start a session"

Search Memory

palace(operation="search", query="why did we switch to GraphQL")

File a Memory

palace(operation="remember", wing="wing_myapp", content="Chose Clerk for auth")
Technical Highlights

Built for production memory

~1:1.2 Test Ratio

4,313 lines of source, 5,347 lines of tests across 5 modules. Integration tests exercise the full palace lifecycle — briefing, capture, interject, and curation.

Module Architecture

  • tool-mempalace — 8 operations: search, remember, kg, traverse, diary, mine, events, garden
  • hooks-mempalace-briefing — session-start wake-up
  • hooks-mempalace-capture — verbatim auto-filing
  • hooks-mempalace-interject — mid-session surfacing
  • hooks-project-context — coordination files

Local-First

All data stays on your machine. Palace storage at ~/.mempalace/. Nothing leaves your environment.

Zero Deletion

Duplicates are preserved with low importance, never dropped. KG enrichment and garden analysis are purely additive.

Graceful Degradation

Without MemPalace installed, the bundle runs in coordination-files-only mode. Palace features silently skip.

The Bigger Picture

Memory as
infrastructure

MemPalace consolidates five previously separate memory modules into a single bundle. It's the persistence layer that makes agent continuity possible.

Consolidation

Supersedes amplifier-bundle-memory, amplifier-bundle-project-memory, amplifier-module-context-memory, amplifier-module-tool-memory, and amplifier-module-hooks-memory-capture into one cohesive system.

Research-Backed

Includes a 14-page research paper covering architecture, event observability design, KG intelligence with formula proofs, evaluation methodology, and philosophy preservation analysis.

The palace metaphor isn't decorative — it gives agents a spatial navigation model. Wings are projects, rooms are topics, drawers are memories. The same mental model humans have used since ancient Greece.

Sources & Methodology

How we built this deck

Primary repository: colombod/amplifier-bundle-memory

Data sources:

Benchmark data:

No metrics were fabricated. All numbers derived from git history and repository documentation as of May 2026.

Get Started

Give your agent
a memory

Install MemPalace and the memory bundle. Your agent will never forget again.

pip install mempalace amplifier bundle add git+https://github.com/colombod/amplifier-bundle-memory@v1.2.1
github.com/colombod/amplifier-bundle-memory
MIT License · Python 3.11+ · Local-first
More Amplifier Stories