Can you compare this vendor against the one I liked last week?
Memory GUMem
Your agent forgets too much.
We fixed that.
Most agents have goldfish memory — they forget the user across sessions. GUMem gives you a memory layer that captures chat + behavior + intent, with white-box inspectability and per-event audit. Every memory write reviewed. Every recall traced.
The choice
Why it exists
Without it, every conversation starts cold. Users repeat their preferences, retell their history, re-explain their context. Worse, the agent never learns from behavior — clicks, searches, and tool calls vanish the moment the session ends. Your "intelligent assistant" is a stateless function.
GUMem gives the agent a real memory. Dual-track: conversation + behavior. Three layers of decay-aware compression: Facts → Summary → Recall. Every write inspectable, every recall traced to source. Your agent actually gets to know your user.
Memory sources
GUMem remembers what users say, and what they actually do.
Chat alone is not memory. GUMem captures the conversation, the behavior around it, and the tool outcomes that reveal intent. The result is context the agent can inspect, govern, and reuse across sessions.
2memory tracks3compression layers∞session continuity
GGUMemwhite-box memory
Chatmessages
Browserclicks + visits
Searchqueries
Emailthreads
CRMaccounts
Toolscalls + results
Recall workflow
A new session can start with the context the user already earned.
GUMem turns past messages and actions into inspectable memory objects, then retrieves the right mix of short-term and long-term context for the next answer.
Facts
User prefers concise answers
Summary
Compares vendors and cares about deployment risk
Recall
Use concise context in the next answer
Context-aware response
Compared against Acme from last week. This vendor is cheaper, but weaker on audit export and deployment controls.
source: chatsource: clickstream
Memory governance
Memory that can be inspected before it influences the model.
GUMem makes memory reviewable by design. Sanitize before write, approve important updates, and trace every recall back to the raw event that created it.
Strip secrets, normalize entities, and block unsafe memory entries.
Keep high-impact propositions visible before they shape future answers.
Show the source event, confidence, and time decay behind each memory.
How it works
Memory production lineWHAT HAPPENED → WHO THEY ARE → CONTEXT
Messages
Action Log
- 01
Facts
Raw events → summarized facts
→ prefers light · size 42 · runs at dawn
- 02
Summary
Facts → inferred user profile
→ morning runner · cushion fan · brand-loyal
- 03
Recall
Profile → formatted context
→ short / mid / long-term
LLM
WebHooksIntercept at any stage. Sanitize, inject rules, sync to your audit pipeline.
before_addSanitize before write
before_llmInject rules into prompt
after_llmSync to audit / CRM
- 1Capture: messages and action logs flow into GUMem as raw events.
- 2Facts: raw events distill into summarized facts — entities, preferences, time ranges.
- 3Summary: facts get inferred into a user profile — themes, traits, confidence scores.
- 4Recall: a query returns short-term + mid-term + long-term context, formatted for the prompt.
- 5WebHooks at every stage let you govern, inject rules, or sync to CRM.
What you can do
Three capabilities. One SDK call away.
Cross-session recall
Open a new session and the agent already knows your user. Short, mid, long-term context — formatted for the LLM, ready to drop into the system prompt.
See it in docs →Dual-track memory
Conversation and behavior captured separately. The agent learns from clicks, filters, and tool calls — not just what users type into the chat box.
See it in docs →WebHook governance
Three stages — sanitize before write, inject rules before the LLM, sync to your audit pipeline after. Govern memory before it ever touches the model.
See it in docs →Code samples and the full API live in the docs. This page tells you why; docs tell you how.