Agent Memory Systems 2026 — Filesystem vs Database

Key Insight: Don't Conflate Interface with Substrate

Filesystems win as an interface (LLMs already know how to use them)

Databases win as a substrate (concurrency, auditability, semantic search)

Why Memory Matters

LLMs are stateless by default — each request starts from scratch. Memory systems:

Context windows reset with each API request. Memory provides long-term recall across sessions.

Memory Types

Short-term (Working Memory)

Immediate context for current task. Maintains state within multi-step queries.

Long-term (Persistent Memory)

Sessions, decisions, patterns, learned behaviors. Survives beyond single interaction.

Vector Retrieval (Semantic Search)

Finds content by meaning, not keywords. Critical as knowledge base grows — grep degrades on paraphrases.

Filesystem vs Database Tradeoffs

When Filesystems Win

Prototypes and single-agent systems:

[REDACTED]'s vault validates this approach — Daily notes, research files, Obsidian sync.

When Databases Win

Shared state and production systems:

Redis + vector search = memory infrastructure layer

Production Patterns

Three-Tier Memory Architecture (Redis pattern)

Agent Coordination

When multiple agents share memory:

Implications for [REDACTED]'s Squad

Filesystem Approach (Current):

Database Approach (Production):

Hybrid Pattern (What Squad Does):

Current Tool Gap

telegram-memory-search currently uses grep — keyword search, not semantic.

Fix: Add vector search for meaning-based retrieval:

Sources

Added: 2026-03-01

Status: Active research — agent memory architectures

Next: Add semantic search to vault-search (vector database integration)