AuthorsJoel Ward
TL;DR
MemoriesDB fuses temporal ordering, semantic embeddings, and graph relations into a single append-only store to maintain long-term coherence for LLM agents without custom infrastructure.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
MemoriesDB targets context decoherence, where long interactions cause facts and intentions to drift out of scope and reasoning fragments into disjoint episodes.
LLM-based agents relying on sliding windows or RAG lose persistent structure, so accumulated knowledge cannot be reasoned about as a coherent whole over hours, days, or weeks.
HOW IT WORKS
MemoriesDB centers on The Memory Record, Edges and Relations, and the Temporal Semantic Stack to unify time, meaning, and connection in one append-only schema.
You can think of MemoriesDB like a laminated card catalog of experiences, where each card has coordinates in time, a semantic vector, and arrows to related cards.
This geometric triality lets MemoriesDB support time-bounded retrieval, hybrid semantic search, and lightweight structural reasoning that a plain context window or standalone vector database cannot provide.
DIAGRAM
This diagram shows how MemoriesDB processes a query by combining temporal filters, semantic similarity, and graph expansion into a single retrieval path.
DIAGRAM
This diagram shows how MemoriesDB’s background maintenance preserves coherence through embedding normalization, edge pruning, and coherence sampling.
PROCESS
The Memory Record
MemoriesDB first encodes each experience as The Memory Record tuple (ti, κi, Vi, mi), assigning microsecond timestamps and multi view embeddings for later retrieval.
Edges and Relations
MemoriesDB then creates Edges and Relations between Memory Records, labeling relations like reply or summary of and storing strengths and confidences as edge weights.
The Temporal Semantic Stack
MemoriesDB organizes Memory Records into the Temporal Semantic Stack, treating each timestamp as a local plane and stacking planes to preserve cross temporal coherence.
Query Semantics
During Query Semantics, MemoriesDB applies temporal windows, semantic similarity, and relational filters in one pipeline to return coherent context for the agent’s response.
KEY CONTRIBUTIONS
Unified data model
MemoriesDB introduces a unified data model where The Memory Record, Edges and Relations, and the Temporal Semantic Stack coexist in a single append only store with pgvector embeddings.
Geometric memory formulation
MemoriesDB models experience as a time indexed graph of temporal semantic surfaces, using distance and coherence metrics to track semantic drift over long horizons.
Working implementation
MemoriesDB ships a working PostgreSQL implementation that supports hybrid vector SQL queries, graph traversal, and coherence tracking on commodity hardware without specialized engines.
RESULTS
Single insert latency
2.5 ms
context: measured at 1 M records on PostgreSQL 16
Batch insert throughput
8000 recs/s
context: batch size 100 at 1 M records
Batch insert peak
10000 recs/s
context: batch size 100 at 100 records
Memory scale
10 M records
context: interactive latency for hybrid queries on medium datasets
The prototype benchmarks on a 32 core workstation with 128 GB RAM illustrate that MemoriesDB can sustain thousands of inserts per second while keeping hybrid temporal semantic graph queries interactive. These observations support the main result that MemoriesDB’s unified architecture is practical on commodity SQL infrastructure.
BENCHMARK
The prototype benchmarks on a 32 core workstation with 128 GB RAM illustrate that MemoriesDB can sustain thousands of inserts per second while keeping hybrid temporal semantic graph queries interactive. These observations support the main result that MemoriesDB’s unified architecture is practical on commodity SQL infrastructure.
BENCHMARK
Insertion latency and throughput for MemoriesDB on PostgreSQL 16 with pgvector.
KEY INSIGHT
MemoriesDB shows that a unified temporal semantic relational memory can run efficiently on plain PostgreSQL 16 with pgvector, without specialized vector or graph databases.
This is surprising because many assume long term coherent memory requires custom engines, but MemoriesDB demonstrates that careful schema design and append only geometry are enough.
WHY IT MATTERS
MemoriesDB gives builders a concrete way to store, retrieve, and reason over long horizon agent experience as a coherent temporal semantic graph.
With MemoriesDB, developers can turn stateless LLMs into continuous learning agents that maintain identity, track semantic drift, and run auto RAG style retrieval directly on their own memory substrate.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…