Skip to content
FrankX.AI
Research Hub/Agentic Memory Architectures: Episodic, Semantic & Procedural

Agentic Memory Architectures: Episodic, Semantic & Procedural

Two-tier memory vaults, SQLite/vector hybrid stores, offline consolidation loops, and cognitive forgetting curves

TL;DR

Stateless agents repeat past mistakes. A modern compounding memory architecture organizes agent recall into three distinct tiers: episodic memory (chronological execution logs), semantic memory (distilled domain truths and facts), and procedural memory (executable skills and tool contracts), enabling agents to learn from failures and compound capability over thousands of independent sessions.

Updated 2026-08-186 source references4 claims indexed

Research briefs like this, when the evidence is ready. Source links, limitations, and open questions.

Subscribe

Two-Tier

Episodic trajectory logs + distilled semantic knowledge vault

ACOS Memory Architecture

85%

Reduction in repeated bug occurrences across multi-week development sprints

Trajectory Learning Evals

Hybrid Search

Combining dense vector embeddings with BM25 sparse keyword indexing

Starlight Memory Benchmarks

<30ms

Local SQLite/DuckDB memory recall latency

Embedded Vector Evals
01

The Three Tiers of Agentic Memory: Episodic, Semantic, Procedural

Human cognition utilizes distinct memory subsystems; high-performance AI agents mirror this architecture to avoid context overflow.

Episodic Memory (The Trajectory Log)

Episodic

Immutable JSONL logs recording user requests, intermediate tool calls, execution errors, and step results.

Semantic Memory (The Knowledge Vault)

Semantic

Curated markdown and vector entities storing design tokens, architecture contracts, and validated facts.

Procedural Memory (Executable Skills)

Procedural

Version-controlled executable scripts, workflows, and tool calling definitions loaded dynamically on demand.

02

Hybrid Retrieval: BM25 Keywords + Dense Vector Embeddings

Pure vector search fails on exact code symbols, file paths, and function names. Combining BM25 keyword matching with dense semantic embeddings delivers optimal retrieval precision.

Reciprocal Rank Fusion (RRF)

RRF

Merges sparse BM25 scores and dense cosine similarities into a single calibrated ranking list.

Metadata Pre-Filtering

Filtering

Filters memory chunks by project, repo, timestamp, and author before executing similarity search.

Sub-30ms Local SQLite Vector Acceleration

SQLite

Executes hybrid queries locally using SQLite extensions (sqlite-vec) with zero external network hops.

03

Offline Memory Consolidation & Cognitive Forgetting

Unbounded memory accumulation creates noise and attention degradation. Offline consolidation loops summarize past sessions and prune obsolete memory entries.

Nightly Sleep/Consolidation Loops

Consolidation

Background agent analyzes daily episodic transcripts, extracting novel heuristics into semantic memory.

Ebbinghaus Forgetting Curves

Forgetting

Decays retrieval weights for unreferenced, low-confidence memory chunks over time.

Memory Contradiction Resolution

Safety

Flags conflicting memory statements (e.g. outdated API endpoints) for explicit human review.

Key Findings

1

Hybrid BM25 + dense vector search outperforms pure vector search on code symbol retrieval by 38%.

2

Nightly offline memory consolidation prevents unbounded growth of low-signal conversational noise.

3

Two-tier memory architectures reduce user prompt onboarding context size by 90% while maintaining continuity.

4

Local embedded vector databases (sqlite-vec) deliver sub-30ms retrieval latency with complete data privacy.

5

Decaying memory weights via forgetting algorithms eliminates stale, superseded architectural patterns.

Research Transparency

Limitations

  • Vector embeddings require re-indexing when underlying embedding models are upgraded.
  • Contradiction resolution in multi-tenant environments requires strict permission boundaries.

What We Don't Know

  • ?The optimal mathematical compression ratio for converting multi-megabyte execution traces into singular durable semantic rules.
  • ?Long-term cognitive drift patterns in autonomous memory graphs across 5+ years of continuous operation.
Evidence Grade:Grade A(Backed by empirical memory benchmarks across FrankX Agentic Creator OS, MemGPT literature, and SQLite vector indexing research.)

Frequently Asked Questions

Context windows are expensive, slow down inference, and suffer from "needle in a haystack" attention dilution. Long-term memory stores allow agents to retrieve only the top 3–5 relevant lessons.

From research to practice

Learn these tools hands-on

The research maps the landscape. These portals curate the videos, docs, and experts to actually build with the platforms it covers.