🧠 A2A Memory System

262x

Graph Queries

4ms vs 1001ms

2.7x

Repeated Queries

67% fewer API calls

0→10

Knowledge Recall

Amnesia → Full memory

1.8x

Cross-Agent

Instant context sharing

8ms

Search Speed

FTS5 per operation

33%

API Saved

600 calls/month

📊 All Tests

⚡ Stress Test

💰 ROI

🎯 Conclusions

🔄 Test 1: Repeated Query ×3

2.7x faster

Customer asks "Who are our competitors?" three times — a typical scenario.

🔴 Without Memory

3,004ms

15 API calls (full research ×3)

🟢 With Memory

1,107ms

5 API calls (research once, cache twice)

Without3,004ms

With Memory1,107ms

💡 Query 2 from graph: 3ms (vs 1,001ms). Query 3 via FTS5: 3ms. Saves 67% of API calls.

🔗 Test 2: Cross-Agent Context

1.8x faster

Researcher finds data, Writer needs it. Without memory — Writer repeats the search.

🔴 Without Memory

1,102ms

Writer wastes 500ms re-searching

🟢 With Memory

625ms

Writer reads 2 observations instantly

Without1,102ms

With Memory625ms

💡 Agents share context through persistent knowledge graph — no duplicate work.

🧠 Test 3: 10 Sessions — Knowledge Recall

0 → 10

After 10 work sessions, ask: "What technologies have we studied?"

🔴 Without Memory

0 knowledge

Complete amnesia — forgets everything

🟢 With Memory

10 entities

Full recall in 4ms + relations + FTS

💡 Knowledge accumulates over time. The longer the customer uses the system, the smarter it gets.

🕸️ Test 4: Graph Impact Analysis

262x faster

Question: "What will Bug #42 in the API service affect?" — needs transitive dependency traversal.

🔴 Without Memory

1,001ms

3 sequential agent queries, incomplete result

🟢 With Memory

4ms

BFS depth=2: 5 entities, 100% coverage

Without1,001ms

With Memory4ms

💡 Graph traversal: Bug→API→DB, API→Auth, Project→API — complete dependency map in a single query.

⚡ Stress Test — 170 Operations on Live System

15.7ms

Write (observe)

100 ops → 1,566ms total

8.0ms

Search (FTS5)

50 ops → 400ms total

9.3ms

Graph (BFS)

20 ops → 186ms total

+16KB

DB Growth

100 observations added

Entities After

From 35 → 55

137

Observations

From 37 → 137

💡 System stable under sustained load. SQLite + FTS5 holds sub-16ms latency. 5MB RAM footprint. Zero errors.

💰 ROI — Monthly API Savings

Model: 20 tasks/day × 30 days = 600 tasks/month. 40% repeat rate. 3 API calls per task.

1,800

API calls without memory

→

1,200

API calls with memory

→

600

calls saved (33%)

💡 At 100 customers: 60,000 API calls/month saved. Memory pays for itself by extending free tier limits.

🕸️

262x

Graph queries are the killer feature. Impact analysis in 4ms vs manual agent polling in 1s.

🔄

2.7x

Repeated queries — the most common case. Customers return to the same topics.

🔗

1.8x

Cross-agent context. Writer produces better output with Researcher's data instantly.

🧠

0→10

Knowledge accumulation. The system gets smarter the longer you use it.

⚡

8-16ms

Production-ready. SQLite + FTS5 on any lightweight VM.

💰

33%

API savings. Memory extends free tier limits at scale.

Technology Stack

SQLite + FTS5 (Full-Text Search) + Knowledge Graph (entities, relations, observations)
18 REST API endpoints · 42/42 tests passing · 5MB RAM · Port 5950
📂 GitHub: a2a-agent-saas/a2a-memory