I Tested 4 Agent Memory SDKs: All Failed on Contradiction Updates and Entity Resolution
Dismal-Account-1151 · reddit · 2026-09-25
Building an agent, the author tested 4 memory SDKs and 4 memory tools with the same scenario: told the agent in session 1 he prefers dark mode, in session 5 that he switched to light mode, then in session 8 asked which mode he uses.
All failed: none removed the old fact when the new one arrived. One just added the new fact next to the old; another stored both separately with nothing linking them; one returned the newer one most of the time but the old one resurfaced with different phrasing; the last required manual edits and doesn't handle contradictions at all. The model simply picks whichever answer matches the query better, so answers vary by phrasing.
Entity resolution was equally bad: mentioning "vansh," "vansh from india," and "VS" produced three separate entries in one tool, two linked but missing the shorthand in another, three different people in a third, and no entity resolution at all in the last.
His advice: give any memory tool a fact that changes across sessions and mention the same person three ways; if it can't handle either, it's just similarity search over old conversations.
More from coding & agent
- Exa launches Agent Ultra, an agent-swarm deep research tool claiming SOTA — yoimnotkesku · 2026-09-26
- YC Backs 'Spend More Compute on Validation': Greptile Launches Plus and Apex Review Tiers — ycombinator · 2026-09-26
- OpenAI Showcases Proaction Using GPT-6 Astra to Build Fleet-Management Agents Faster — OpenAIDevs · 2026-09-26
- Prompt-to-Promo Video: Deterministic Frame Rendering Is the Real Unlock — ezshine · 2026-09-26
- Open-source MCP server lets agents use secrets they can never see — JuanixVentures · 2026-09-26
- One Stripe Engineer Merged 600 AI-Written Changes in Six Months, One Rollback — victor_explore · 2026-09-26