Memory OFF
Generic copilot
Waiting for an alert.
Incident response · powered by Hindsight agent memory
Your on-call agentthat never forgets.
The fix for tonight's outage is already in last month's postmortem. RecallOps dives into your incident memory, surfaces what worked (and what failed), and hands the engineer a ranked plan, in seconds.
the problem
At 3 AM, nobody remembers last month's outage.
3×
The same payments service broke after a deploy three times in eight weeks, with three different root causes.
20 min
Lost each time someone scaled the pods again. Two postmortems already said scaling did nothing.
0
Things a generic LLM copilot knows about your history. It gives the textbook checklist, every time.
live triage
Same alert. Same model. One has memory.
Pick an alert. The left agent answers from nothing. The right one calls Hindsight recall + reflect over ShopNova's incident history (synthetic) before planning.
0/500
Memory ON
RecallOps + Hindsight
Waiting for an alert.
Recalled from the deep
Glowing nodes appear here: one per past incident Hindsight surfaces.
Reflect
Hindsight reasons over the whole bank, guided by an SRE mission: cite incidents, prefer what worked, warn about what failed.
learning curve
Dive deeper: watch it get smarter.
You get a private, empty memory bank. Interaction 1 knows nothing. Retain one postmortem, ask the same question, and the plan goes from guesswork to specific.
SCENARIO ALERT
- 01Interaction 1 · empty memory
- 02Engineer resolves it · retain
- 03Interaction 2 · same alert
The postmortem that gets retained (synthetic)
Before · interaction 1
—
After · interaction 2
—
retain
Every resolved incident sinks into memory.
Write down what fixed it and what didn't. It is retained into the shared ShopNova bank with its timestamp, service and team tags, and the next similar alert benefits.
the memory floor
What the agent actually remembers.
Facts and consolidated observations that Hindsight extracted from postmortems, laid out on a timeline. Ask it anything.
how it works
Three calls. One memory bank.
retain()
Each postmortem goes in with its real incident timestamp, the incident ID as document_id, and tags like service:payments-api. Hindsight extracts facts and entities and links them.
recall()
Semantic + keyword + graph + temporal retrieval, reranked and cut to a 2,048-token budget. The top facts, with source incident IDs, ground the LLM.
reflect()
An agentic loop over the whole bank with an SRE mission: cite incidents, prefer fixes that worked, warn about fixes that failed.