The Agent That Remembers

A five-tier persistent memory for Amplifier

The Problem & The Build

A brilliant colleague with permanent amnesia

Without persistent memory an Amplifier agent starts every session from zero, and you re-explain everything you already told it. amplifier-bundle-memory — originally called MemPalace, a name retired in the v2.0.0 native cutover — gives the agent a persistent five-tier memory that carries across sessions.

First question: can it actually find the right memory when it needs it?

Proof — Retrieval

The right memory is almost always findable

The retrieval engine hits 96.6% Recall@5 on LongMemEval — raw, with no LLM rerank. Held-out and reranked runs push it higher still.

So the substrate works. Now: how is the memory actually organized?

96.6%

R@5 — LongMemEval (raw, no LLM rerank)

98.4%

R@5 — LongMemEval (hybrid v4, held-out)

88.9%

R@10 — LoCoMo (hybrid v5, top-10)

The Architecture

The memory is five tiers, boot-time to background

Layered from the context loaded at startup down to silent verbatim filing during the session — each tier does one job.

One tier surfaces memories mid-session. That is where it gets hard.

Tier 0IdentityCoordination files into the system prompt

Tier 1BriefingEphemeral briefing at session:start

Tier 2InterjectionGated injection mid-session

Tier 3On-DemandArchivist multi-hop KG traversal

+Capture & CurationSilent verbatim filing, dedup at session:end

The Complication

Findable isn't the same as useful

Surface every relevant memory mid-session and you flood the context window. The obvious fix — inject everything that matches — breaks the very thing it is meant to help.

Unprompted recall is a precision problem, not a volume one. So how do you engineer restraint?

Inject everything relevant, and you help nothing.

The cost of a wrong interruption is a polluted context window.

Mechanism — The Gate

Tier 2 only speaks when it is sure

The Interject hook decides whether a memory is worth interrupting for using a single similarity gate — with a fallback judge for the uncertain middle band.

Restraint decides when to speak. The next choice keeps what it remembers accurate.

1

Cosine similarity gateInject only above threshold 0.72

2

LLM-as-judge fallbackFor the uncertain 0.62–0.72 band

3

Ephemeral system hintSurfaced mid-session, unprompted, at the right moment

Mechanism — Storage

Store it verbatim, never summarize it

Content is filed as-is and never compressed into a summary. The raw text is exactly what keeps retrieval sharp — no lossy summarization, no drift.

Precise gate plus verbatim storage. Now watch what that restraint actually buys.

Verbatim storage

Content is stored as-is, never summarized

Why it matters

"This is what makes retrieval accurate at 96.6% R@5" — context/instructions.md

The Payoff

Restraint is the feature, not a limitation

In the LongMemEval-S study the Interject hook rejected 245 candidate memories and fired just 4 times — an 8% injection rate. Yet those 4 injections lifted the LLM Judge score from 4.0% to 10.0%.

Saying nothing 92% of the time is exactly what makes the memory trustworthy.

4

Injections fired (8% rate)

245

Candidate memories rejected

4.0% → 10.0%

LLM Judge score, baseline (no memory) → with Interject

The Takeaway

Memory becomes infrastructure

Continuity arrives at session start, the right memory surfaces only when it matters, and you stop re-explaining what you already told the agent. It runs local-first — nothing leaves your machine by default via the local embedding model.

Reliable, restrained, local: memory the agent keeps so you don't have to.

Sources

Research Methodology

Active — v2.0.0 native cutover complete

Subject: michaeljabbour/amplifier-bundle-memory (tool now named "memory"; "MemPalace" was the original name, removed in the v2.0.0 native cutover, 2026-07-07).

Data as of: most recent commit 2026-07-08; independently re-verified against a live clone at /tmp/mem-primary.

Research performed:

Retrieval metrics: grep -rni '96.6' README.md docs/eval/EVALUATION.md skills/memory/SKILL.md context/instructions.md (README.md:236; 96.6% / 98.4% / 88.9%).
Five-tier architecture: grep -rni 'five-tier' docs/eval/EVALUATION.md then sed -n '22,33p' docs/eval/EVALUATION.md.
Interject gate & injection study: sed -n '30p;46p;63p;87p' docs/eval/EVALUATION.md (0.72 cosine gate; 0.62–0.72 LLM-judge band; 245 rejected / 4 injected / 8%; 4.0% → 10.0%).
Verbatim storage: grep -ni 'verbatim' context/instructions.md README.md (context/instructions.md:39).
Contributors & timeline: git log --format='%an' | sort | uniq -c, git rev-list --count HEAD (55 commits), gh pr list --repo michaeljabbour/amplifier-bundle-memory --state all.

Primary contributors: Michael J. Jabbour (46 commits, primary author); Diego Colombo (6 commits).

Gaps & care: 96.6% is the retrieval-engine R@5 metric, NOT the frequency of unprompted interjections — the Tier 2 study numbers (8% injection rate; 4.0% → 10.0% LLM Judge) are a distinct LongMemEval-S measurement and are kept separate here. The docs/research PDF was present but not read line-by-line.