Research Preview

Your AI agents remember everything equally.
What if they knew what mattered?

Meaning Memory is a significance-weighted curation layer for AI agent memory. It learns what's important, compiles it under budget, and shows you why. Field-tested. Stage 1 commercial beta in 2026.

$ pip install meaning-memory

What We Found

Existing memory systems are excellent at storage and retrieval. But they treat all memories equally. We studied what happens when you add a significance layer on top. Here's what the data shows.

58%

Better Recall Under Pressure

When context budgets are tight (100 lines), significance-weighted compilation retains more critical facts than relevance ranking alone (54.8%). The curation layer matters most when space is scarce.

100%

Nothing Gets Lost

Every observation stays in the ledger with full provenance. The compiler decides what's top-of-mind, but nothing is permanently discarded. You can always trace back.

+17%

Smarter Extraction

Graph-enriched entity awareness during ingestion improved fact coverage by 17%. The value isn't in graph queries at search time; it's in knowing what to extract.

2 Tiers

Use Standalone or on Top

Run with zero dependencies (SQLite, 30 seconds). Or add significance on top of Mem0, Qdrant, or pgvector. Your existing memory backend stays; we add the curation layer.

Read the paper: The Symbiotic Studio (SSRN) →

How It Works

  1. Observe Your agents generate observations during sessions. Facts, preferences, decisions, lessons, episodes.
  2. Score Each memory gets a significance score. Starts with LLM calibration, improves with human feedback. Not all memories are equal.
  3. Compile At session start, the compiler builds a budget-capped snapshot: the most significant facts first, deterministic, reproducible, auditable.

What Meaning Memory Adds

Existing memory systems handle storage and retrieval well. Meaning Memory is the curation layer that sits on top.

Layer Your Existing Stack + Meaning Memory
What to keep Everything (or recency-based) Significance-weighted curation
Context assembly Best-effort top-K Deterministic budget compilation
Audit trail Metadata timestamps Full provenance (what changed, when, why)
Multi-agent memory Shared storage (same view for all) Per-agent significance (same facts, different priorities)
Human feedback Binary remember/forget Gradient scoring that trains the system
Deployment Vendor-specific backend Standalone (SQLite) or on top of Mem0 / Qdrant / pgvector

Request Early Access

Stage 1 NDA design partners now. Stage 2 commercial beta planned for late 2026. Get notified and shape the roadmap.