asanisima therobot.is

An entry point to therobot.is

Breathe life into your agent.

Most agents begin every session with no record of the last one. therobot.is gives an agent a memory it keeps between sessions, recall that weighs where and when something happened as well as what was said, and one identity across the interfaces it works through.

Learn more at therobot.is

A memory record at the centre, linked to five neighbours (one per recall signal) and to twelve further records. Line thickness shows association strength.
Fig. 1 One record and its associations. Inner links are coloured by recall signal; line weight is association strength.

What the agent keeps between sessions

  1. 01

    Memory records

    Each observation is stored with its content, its context (where and when it happened) and an affect estimate. Records link to the people, agents, projects and artifacts they mention, so the store is a graph rather than a list.

  2. 02

    Recall across five signals

    A query is ranked separately by semantic, lexical, contextual, affective and structural similarity. The graph walk starts from the semantic and affective results, so recall returns records connected to the query, not only records that resemble it.

  3. 03

    Grading for the one remembering

    After the rankings are fused, results are re-graded for the current task and for recency, and shaped to the persona doing the recalling: the set most useful to this agent, on this task, now.

  4. 04

    One identity, several interfaces

    The same memory and persona serve the agent in a terminal, a chat, a voice call or a browser. Access limits apply twice: inside the query, and again as a filter after fusion.

How five rankings become one

Each signal produces its own ordered list. The lists are combined by reciprocal rank fusion: a record scores the sum, over every list it appears in, of its signal's weight divided by sixty plus its rank in that list.

score(m) = Σp wp / (k + rankp(m)), k = 60

A record that ranks moderately on several signals can outrank one that ranks first on a single signal. Figure 2 shows this: m-238 is first on two lists and present on all five, so it leads the fused list.

How associations strengthen

When two records are retrieved together, the system treats the co-retrieval as evidence that they are related. It creates an edge between them, or strengthens an existing one, in proportion to how often the co-retrieval recurs: a Hebbian update applied to retrieval events.

sem

  1. m-238
  2. m-117
  3. m-041
  4. m-302
  5. m-189

lex

  1. m-041
  2. m-238
  3. m-013
  4. m-255
  5. m-117

ctx

  1. m-117
  2. m-302
  3. m-238
  4. m-076
  5. m-041

aff

  1. m-238
  2. m-076
  3. m-189
  4. m-117
  5. m-013

str

  1. m-302
  2. m-238
  3. m-117
  4. m-255
  5. m-076

fused

  1. m-2380.0809
  2. m-1170.0794
  3. m-3020.0481
  4. m-0410.0477
  5. m-0760.0471
Fig. 2 Five top-five rankings fused with equal weights, k = 60. Illustrative data.

The memory service and the agent suite built around it are documented at therobot.is.

Learn more at therobot.is